# partiallysupervised_image_captioning__e1b880ff.pdf

Partially-Supervised Image Captioning

Peter Anderson Macquarie University Sydney, Australia p.anderson@mq.edu.au

Stephen Gould Australian National University Canberra, Australia stephen.gould@anu.edu.au

Mark Johnson Macquarie University Sydney, Australia mark.johnson@mq.edu.au

Image captioning models are becoming increasingly successful at describing the content of images in restricted domains. However, if these models are to function in the wild for example, as assistants for people with impaired vision a much larger number and variety of visual concepts must be understood. To address this problem, we teach image captioning models new visual concepts from labeled images and object detection datasets. Since image labels and object classes can be interpreted as partial captions, we formulate this problem as learning from partiallyspeciﬁed sequence data. We then propose a novel algorithm for training sequence models, such as recurrent neural networks, on partially-speciﬁed sequences which we represent using ﬁnite state automata. In the context of image captioning, our method lifts the restriction that previously required image captioning models to be trained on paired image-sentence corpora only, or otherwise required specialized model architectures to take advantage of alternative data modalities. Applying our approach to an existing neural captioning model, we achieve state of the art results on the novel object captioning task using the COCO dataset. We further show that we can train a captioning model to describe new visual concepts from the Open Images dataset while maintaining competitive COCO evaluation scores.

1 Introduction

The task of automatically generating image descriptions, i.e., image captioning [1 3], is a longstanding and challenging problem in artiﬁcial intelligence that demands both visual and linguistic understanding. To be successful, captioning models must be able to identify and describe in natural language the most salient elements of an image, such as the objects present and their attributes, as well as the spatial and semantic relationships between objects [3]. The recent resurgence of interest in this task has been driven in part by the development of new and larger benchmark datasets such as Flickr 8K [4], Flickr 30K [5] and COCO Captions [6]. However, even the largest of these datasets, COCO Captions, is still based on a relatively small set of 91 underlying object classes. As a result, despite continual improvements to image captioning models and ever-improving COCO caption evaluation scores [7 10], captioning models trained on these datasets fail to generalize to images in the wild [11]. This limitation severely hinders the use of these models in real applications, for example as assistants for people with impaired vision [12].

In this work, we use weakly-annotated data (readily available in object detection datasets and labeled image datasets) to improve image captioning models by increasing the number and variety of visual concepts that can be successfully described. Compared to image captioning datasets such as COCO Captions, several existing object detection datasets [14] and labeled image datasets [15, 16] are much larger and contain many more visual concepts. For example, the recently released Open Images dataset V4 [14] contains 1.9M images human-annotated with object bounding boxes for 600 object

Now at Georgia Tech (peter.anderson@gatech.edu)

32nd Conference on Neural Information Processing Systems (Neur IPS 2018), Montréal, Canada.

FSA construction

Step 1: Caption completion

Step 2: Parameter update

A man on a scooter riding down the road in front of some other traffic.

CBS decoding

(using current model parameters)

person, bus, scooter,

van, white, yellow

man on traffic . <EOS>

<SOS> A man other traffic .

Figure 1: Conceptual overview of partially-speciﬁed sequence supervision (PS3) applied to image captioning. In Step 1 we construct ﬁnite state automata (FSA) to represent image captions partiallyspeciﬁed by object annotations, and use constrained beam search (CBS) decoding [13] to ﬁnd high probability captions that are accepted by the FSA. In Step 2, we update the model parameters using the completed sequences as a training targets.

classes, compared to the 165K images and 91 underlying object classes in COCO Captions. This reﬂects the observation that, in general, object detection datasets may be easier to scale possibly semi-automatically [17, 18] to new concepts than image caption datasets. Therefore, in order to build more useful captioning models, ﬁnding ways to assimilate information from these other data modalities is of paramount importance.

To train image captioning models on object detections and labeled images, we formulate the problem as learning from partially-speciﬁed sequence data. For example, we might interpret an image labeled with scooter as a partial caption containing the word scooter and an unknown number of other missing words, which when combined with scooter in the correct order constitute the complete sequence. If an image is annotated with the object class person , this may be interpreted to suggest that the complete caption description must mention person . However, we may also wish to consider complete captions that reference the person using alternative words that are appropriate to speciﬁc image contexts such as man , woman , cowboy or biker . Therefore, we characterize our uncertainty about the complete sequence by representing each partially-speciﬁed sequence as a ﬁnite state automaton (FSA) that encodes which sequences are consistent with the observed data. FSA are widely used in natural language processing because of their ﬂexibility and expressiveness, and because there are well-known techniques for constructing and manipulating such automata (e.g., regular expressions can be compiled into FSA).

Given training data where the captions are either complete sequences or FSA representing partiallyspeciﬁed sequences, we propose a novel two-step algorithm inspired by expectation maximization (EM) [19, 20] to learn the parameters of a sequence model such as a recurrent neural network (RNN) which we will use to generate complete sequences at test time. As illustrated in Figure 1, in the ﬁrst step we use constrained beam search decoding [13] to ﬁnd high probability complete sequences that satisfy the FSA. In the second step, we learn or update the model parameters using the completed dataset. We dub this approach PS3, for partially-speciﬁed sequence supervision. In the context of image captioning, PS3 allows us to train captioning models jointly over both image caption and object detection datasets. Our method thus lifts the restriction that previously required image captioning models to be trained on paired image-sentence corpora only, or otherwise required specialized model architectures to be used in order to take advantage of other data modalities [21 24].

Consistent with previous work [13, 21 24], we evaluate our approach on the COCO novel object captioning splits in which all mentions of eight selected object classes have been eliminated from the caption training data. Applying PS3 to an existing open source neural captioning model [10], and

s0 start s1 s2 s3 a Σ c

(a) A sequence (a, <unk>, c) from vocabulary Σ where the length of the missing subsequence <unk> is unknown.

s0 start s1 s2 the score

(b) A sequence that doesn t mention the score .

s0 start s1 s2 s3

s4 s5 s6 s7

D2 D2 D2 D2

(c) A sequence that mentions word(s) from at least two of the three disjunctive sets D1,D2 and D3.

Figure 2: PS3 is a general approach to training RNNs on partially-speciﬁed sequences. Here we illustrate some examples of partially-speciﬁed sequences that can be represented with ﬁnite state automata. Unlabeled edges indicate default transitions , i.e., an unlabeled edge leaving a node n is implicitly labeled with Σ \ S, where S is the set of symbols on labeled edges leaving n and Σ is the complete vocabulary.

training on auxiliary data consisting of either image labels or object annotations, we achieve state of the art results on this task. Furthermore, we conduct experiments training on the Open Images dataset, demonstrating that using our method a captioning model can be trained to identify new visual concepts from the Open Images dataset while maintaining competitive COCO evaluation scores.

Our main contributions are threefold. First, we propose PS3, a novel algorithm for training sequence models such as RNNs on partially-speciﬁed sequences represented by FSA (which includes sequences with missing words as a special case). Second, we apply our approach to the problem of training image captioning models from object detection and labeled image datasets, enabling arbitrary image captioning models to be trained on these datasets for the ﬁrst time. Third, we achieve state of the art results for novel object captioning, and further demonstrate the application of our approach to the Open Images dataset. To encourage future work, we have released our code and trained models via the project website2. As illustrated by the examples in Figure 2, PS3 is a general approach to training sequence models that may be applicable to various other problem domains with partially-speciﬁed training sequences.

2 Related work

Image captioning The problem of image captioning has been intensively studied. More recent approaches typically combine a pretrained Convolutional Neural Network (CNN) image encoder with a Recurrent Neural Network (RNN) decoder that is trained to predict the next output word, conditioned on the previous output words and the image [1, 25 28], optionally using visual attention [2, 7 10]. Like other sequence-based neural networks [29 32], these models are typically decoded by searching over output sequences either greedily or using beam search. As outlined in Section 3, our proposed partially-supervised training algorithm is applicable to this entire class of sequence models.

Novel object captioning A number of previous works have studied the problem of captioning images containing novel objects (i.e., objects not present in training captions) by learning from image labels. Many of the proposed approaches have been architectural in nature. The Deep Compositional Captioner (DCC) [21] and the Novel Object Captioner (NOC) [22] both decompose the captioning model into separate visual and textual pipelines. The visual pipeline consists of a CNN image classiﬁer that is trained to predict words that are relevant to an image, including the novel objects. The textual pipeline is a RNN trained on language data to estimate probabilities over word sequences. Each pipeline is pre-trained separately, then ﬁne-tuned jointly using the available image and caption data. More recently, approaches based on constrained beam search [13], word copying [33] and

2www.panderson.me/constrained-beam-search

neural slot-ﬁlling [24] have been proposed to incorporate novel word predictions from an image classiﬁer into the output of a captioning model. In contrast to the specialized architectures previously proposed for handling novel objects [21 24], we present a general approach to training sequence models on partially-speciﬁed data that uses constrained beam search [13] as a subroutine.

Sequence learning with partial supervision Many previous works on semi-supervised sequence learning focus on using unlabeled sequence data to improve learning, for example by pre-training RNNs [34, 35] or word embeddings [36, 37] on large text corpora. Instead, we focus on the scenario in which the sequences are incomplete or only partially-speciﬁed, which occurs in many practical applications ranging from speech recognition [38] to healthcare [39]. To the best of our knowledge we are the ﬁrst to consider using ﬁnite state automata as a new way of representing labels that strictly generalizes both complete and partially-speciﬁed sequences.

3 Partially-speciﬁed sequence supervision (PS3)

In this section, we describe how partially-speciﬁed data can be incorporated into the training of a sequence prediction model. We assume a model parameterized by θ that represents the distribution over complete output sequences y = (y1, . . . , y T ), y Y as a product of conditional distributions:

t=1 pθ(yt | y1:t 1) (1)

where each yt is a word or other token from vocabulary Σ. This model family includes recurrent neural networks (RNNs) and auto-regressive convolutional neural networks (CNNs) [29] with application to tasks such as language modeling [30], machine translation [31, 32], and image captioning [1 3]. We further assume that we have a dataset of partially-speciﬁed training sequences X = {x0, . . . , xm}, and we propose an algorithm that simultaneously estimates the parameters of the model θ and the complete sequence data Y .

3.1 Finite state automaton speciﬁcation for partial sequences

Traditionally partially-speciﬁed data X is characterized as incomplete data containing missing values [19, 40], i.e., some sequence elements are replaced by an unknown word symbol <unk>. However, this formulation is insufﬁciently ﬂexible for our application, so we propose a more general representation that encompasses missing values as a special case. We represent each partiallyspeciﬁed sequence xi X with a ﬁnite state automaton (FSA) Ai that recognizes sequences that are consistent with the observed partial information. Formally, Ai = (Σ, Si, si 0, δi, F i) where Σ is the model vocabulary, Si is the set of automaton states, si 0 Si is the initial state, δi : Si Σ Si is the state-transition function that maps states and words to states, and F i Si is the set of ﬁnal or accepting states [41].

As illustrated in Figure 2, this approach can encode very expressive uncertainties about the partiallyspeciﬁed sequence. For example, we can allow for missing subsequences of unknown or bounded length, negative information, and observed constraints in the form of conjunctions of disjunctions or partial orderings. Given this ﬂexibility, from a modeling perspective the key challenge in implementing the proposed approach will be determining the appropriate FSA to encode the observed partial information. We discuss this further from the perspective of image captioning in Section 4.

3.2 Training algorithm

We now present a high level speciﬁcation of the proposed PS3 training algorithm. Given a dataset of partially-speciﬁed training sequences X and current model parameters θ, then iteratively perform the following two steps:

Step 1. Estimate the complete data Y by setting yi argmaxy pθ(y | Ai) for all xi X Step 2. Learn the model parameters by setting θ argmaxθ P y Y log pθ(y)

Step 1 can be skipped for complete sequences, but for partially-speciﬁed sequences Step 1 requires us to ﬁnd the most likely output sequence that satisﬁes the constraints speciﬁed by an FSA. As it is typically computationally infeasible to solve this problem exactly, we use constrained beam

Algorithm 1 Beam search decoding

1: procedure BS(Θ, b, T, Σ) With beam size b and vocabulary Σ 2: B {ϵ} ϵ is the empty string 3: for t = 1, . . . , T do 4: E {(y, w) | y B, w Σ} All one-word extensions of sequences in B 5: B argmax E E,|E |=b P

y E Θ(y) The b most probable extensions in E

6: return argmaxy B Θ(y) The most probable sequence

Algorithm 2 Constrained beam search decoding [13]

1: procedure CBS(Θ, b, T, A = (Σ, S, s0, δ, F)) With ﬁnite state recognizer A 2: for s S do 3: Bs {ϵ} if s = s0 else Each state s has a beam Bs

4: for t = 1, . . . , T do 5: for s S do Extend sequences through state-transition function δ 6: Es s S{(y, w) | y Bs , w Σ, δ(s , w) = s} 7: Bs argmax E Es,|E |=b P

y E Θ(y) The b most probable extensions in Es

8: return argmaxy S

s F Bs Θ(y) The most probable accepted sequence

search [13] to ﬁnd an approximate solution. In Algorithms 1 and 2 we provide an overview of the constrained beam search algorithm, contrasting it with beam search [42]. Both algorithms take as inputs a scoring function which we deﬁne by Θ(y) = log pθ(y), a beam size b, the maximum sequence length T and the model vocabulary Σ. However, the constrained beam search algorithm additionally takes a ﬁnite state recognizer A as input, and guarantees that the sequence returned will be accepted by the recognizer. Refer to Anderson et al. [13] for a more complete description of constrained beam search. We also note that other variations of constrained beam search decoding have been proposed [43 45]; we leave it to future work to determine if they could be used here.

Online version The PS3 training algorithm, as presented so far, is inherently ofﬂine. It requires multiple iterations through the training data, which can become impractical with large models and datasets. However, our approach can be adapted to an online implementation. For example, when training neural networks, Steps 1 and 2 can be performed for each minibatch, such that Step 1 estimates the complete data for the current minibatch Y Y , and Step 2 performs a gradient update based on Y . In terms of implementation, Steps 1 and 2 can be implemented in separate networks with tied weights, or in a single network by backpropagating through the resulting search tree in the manner of Wiseman and Rush [46]. In our GPU-based implementation, we use separate networks with tied weights. This is more memory efﬁcient when the number of beams b and the number of states |S| is large, because performing the backward pass in the smaller Step 2 network means that it is not necessary to maintain the full unrolled history of the search tree in memory.

Computational complexity Compared to training on complete sequence data, PS3 performs additional computation to ﬁnd a high-probability complete sequence for each partial sequence speciﬁed by an FSA. Because constrained beam search maintains a beam of b sequences for each FSA state, this cost is given by |S| b γ, where |S| is the number of FSA states, b is the beam size parameter, and γ is the computational cost of a single forward pass through an unrolled recurrent neural network (e.g., the cost of decoding a single sequence). Although the computational cost of training increases linearly with the number of FSA states, for any particular application FSA construction is a modeling choice and there are many existing FSA compression and state reduction methods available.

4 Application to image captioning

In this section, we describe how image captioning models can be trained on object annotations and image tags by interpreting these annotations as partially-speciﬁed caption sequences.

Captioning model For image captioning experiments we use the open source bottom-up and topdown attention captioning model [10], which we refer to as Up-Down. This model belongs to the class of encoder-decoder neural architectures and recently achieved state of the art results on the COCO test server [6]. The input to the model is an image, I. The encoder part of the model consists of a Faster R-CNN object detector [47] based on the Res Net-101 CNN [48] that has been pre-trained on the Visual Genome dataset [49]. Following the methodology in Anderson et al. [10], the image I is encoded as a set of image feature vectors, V = {v1, . . . , vk}, vi RD, where each vector vi is associated with an image bounding box. The decoder part of the model consists of a 2-layer Long Short-Term Memory (LSTM) network [50] combined with a soft visual attention mechanism [2]. At each timestep t during decoding, the decoder takes as input an encoding of the previously generated word given by WeΠt, where We RM |Σ| is a word embedding matrix for a vocabulary Σ with embedding size M, and Πt is one-hot encoding of the input word at timestep t. The model outputs a conditional distribution over the next word output given by p(yt | y1:t 1) = softmax (Wpht + bp), where ht RN is the LSTM output and Wp R|Σ| N and bp R|Σ| are learned weights and biases. The decoder represents the distribution over complete output sequences using Equation 1.

Finite state automaton construction To train image captioning models on datasets of object detections and labeled images, we construct ﬁnite state automata as follows. At each training iteration we select three labels at random from the labels assigned to each image. Each of the three selected labels is mapped to a disjunctive set Di containing every word in the vocabulary Σ that shares the same word stem. For example, the label bike maps to { bike, bikes, biked, biking }. This gives the captioning model the freedom to choose word forms. As the selected image labels may include redundant synonyms such as bike and bicycle, we only enforce that the generated caption mentions at least two of the three selected image labels. We therefore construct a ﬁnite state automaton that accepts strings that contain at least one word from at least two of the disjunctive sets. As illustrated in Figure 2(c), the resulting FSA contains eight states (although the four accepting states could be collapsed into one). In initial experiments we investigated several variations of this simple construction approach (e.g., randomly selecting two or four labels, or requiring more or fewer of the selected labels to be mentioned in the caption). These alternatives performed slightly worse than the approach described above. However, we leave a detailed investigation of more sophisticated methods for constructing ﬁnite state automata encoding observed partial information to future work.

Out-of-vocabulary words One practical consideration when training image captioning models on datasets of object detections and labeled images is the presence of out-of-vocabulary words. The constrained decoding in Step 1 can only produce ﬂuent sentences if the model can leverage some side information about the out-of-vocabulary words. To address this problem, we take the same approach as Anderson et al. [13], adding pre-trained word embeddings to both the input and output layers of the decoder. Speciﬁcally, we initialize We with pretrained word embeddings, and add an additional output layer such that vt = tanh (Wpht + bp) and p(yt | y1:t 1) = softmax (W T e vt). For the word embeddings, we concatenate Glo Ve [37] and dependency-based [51] embeddings, as we ﬁnd that the resulting combination of semantic and functional context improves the ﬂuency of the constrained captions compared to using either embedding on its own.

Implementation details In all experiments we initialize the model by training on the available image-caption dataset following the cross-entropy loss training scheme in the Up-Down paper [10], and keeping pre-trained word embeddings ﬁxed. When training on image labels, we use the online version of our proposed training algorithm, constructing each minibatch of 100 with an equal number of complete and partially-speciﬁed training examples. We use SGD with an initial learning rate of 0.001, decayed to zero over 5K iterations, with a lower learning rate for the pre-trained word embeddings. In beam search and constrained beam search decoding we use a beam size of 5. Training (after initialization) takes around 8 hours using two Titan X GPUs.

5 Experiments

5.1 COCO novel object captioning

Dataset splits To evaluate our proposed approach, we use the COCO 2014 captions dataset [52] containing 83K training images and 41K validation images, each labeled with ﬁve human-annotated

Table 1: Impact of training and decoding with image labels on COCO novel object captioning validation set scores. All experiments use the same ﬁnite state automaton construction. On out-ofdomain images, imposing label constraints during training using PS3 always improves the model (row 3 vs. 1, 4 vs. 2, 6 vs. 5), and constrained beam search (CBS) decoding is no longer necessary (row 4 vs. 3). The model trained using PS3 and decoded with standard beam search (row 3) is closest to the performance of the model trained with the full set of image captions (row 7).

Training PS3 CBS Out-of-Domain Scores In-Domain Scores

Captions Labels Labels SPICE METEOR CIDEr F1 SPICE METEOR CIDEr

1 G # 14.4 22.1 69.5 0.0 19.9 26.5 108.6 2 G # 15.9 23.1 74.8 26.9 19.7 26.2 102.4 3 G # 18.3 25.5 94.3 63.4 18.9 25.9 101.2 4 G # 18.2 25.2 92.5 62.4 19.1 25.9 99.5

5 G # 18.0 24.5 82.5 30.4 22.3 27.9 109.7 6 G # 20.1 26.4 95.5 65.0 21.7 27.5 106.6 7 20.1 27.0 111.5 69.0 20.0 26.7 109.5

= full training set, G # = impoverished training set, = constrained beam search (CBS) decoding with predicted labels, = CBS decoding with ground-truth labels

Table 2: Performance on the COCO novel object captioning test set. + CBS indicates that a model was decoded using constrained beam search [13] to force the inclusion of image labels predicted by an external model. On standard caption metrics, our generic training algorithm (PS3) applied to the Up-Down [10] model outperforms all prior work.

Out-of-Domain Scores In-Domain Scores

Model CNN SPICE METEOR CIDEr F1 SPICE METEOR CIDEr

DCC [21] VGG-16 13.4 21.0 59.1 39.8 15.9 23.0 77.2 NOC [22] VGG-16 - 21.3 - 48.8 - - - C-LSTM [23] VGG-16 - 23.0 - 55.7 - - - LRCN + CBS [13] VGG-16 15.9 23.3 77.9 54.0 18.0 24.5 86.3 LRCN + CBS [13] Res-50 16.4 23.6 77.6 53.3 18.4 24.9 88.0 NBT [24] VGG-16 15.7 22.8 77.0 48.5 17.5 24.3 87.4 NBT + CBS [24] Res-101 17.4 24.1 86.0 70.3 18.0 25.0 92.1 PS3 (ours) Res-101 17.9 25.4 94.5 63.0 19.0 25.9 101.1

captions. We use the splits proposed by Hendricks et al. [21] for novel object captioning, in which all images with captions that mention one of eight selected objects (including synonyms and plural forms) are removed from the caption training set, which is reduced to 70K images. The original COCO validation set is split 50% for validation and 50% for testing. As such, models are required to caption images containing objects that are not present in the available image-caption training data. For analysis, we further divide the test and validation sets into their in-domain and out-of-domain components. Any test or validation image with a reference caption that mentions a held-out object is considered to be out-of-domain. The held-out objects classes selected by Hendricks et al. [21], are BOTTLE, BUS, COUCH, MICROWAVE, PIZZA, RACKET, SUITCASE, and ZEBRA.

Image labels As with zero-shot learning [53], novel object captioning requires auxiliary information in order to successfully caption images containing novel objects. In the experimental procedure proposed by Hendricks et al. [21] and followed by others [13, 22, 23], this auxiliary information is provided in the form of image labels corresponding to the 471 most common adjective, verb and noun base word forms extracted from the held-out training captions. Because these labels are extracted from captions, there are no false positives, i.e., all of the image labels are salient to captioning. However, the task is still challenging as the labels are pooled across ﬁve captions per image, with the number of labels per image ranging from 1 to 27 with a mean of 12.

Evaluation To evaluate caption quality, we use SPICE [54], CIDEr [55] and METEOR [56]. We also report the F1 metric for evaluating mentions of the held-out objects. The ground truth for an object mention is considered to be positive if the held-out object is mentioned in any reference

zebra bus couch microwave

Baseline: A close up of a giraffe with its head. Baseline: A food truck parked on the side of a road.

Baseline: A living room ﬁlled with lots of furniture.

Baseline: A picture of an oven in a kitchen.

Ours: A couple of zebra standing next to each other.

Ours: A white bus driving down a city street. Ours: A brown couch sitting in a living room. Ours: A microwave sitting on top of a counter.

pizza racket suitcase bottle

Baseline: A collage of four pictures of food. Baseline: A young girl is standing in the tennis court.

Baseline: A group of people walking down a street.

Baseline: A woman in the kitchen with a toothbrush in her hand.

Ours: A set of pictures showing a slice of pizza. Ours: A little girl holding a tennis racket. Ours: A group of people walking down a city street.

Ours: A woman wearing a blue tie holding a yellow toothbrush.

Figure 3: Examples of generated captions for images containing novel objects. The baseline Up Down [10] captioning model performs poorly on images containing object classes not seen in the available image-caption training data (top). Incorporating image labels for these object classes into training using PS3 allows the same model to produce ﬂuent captions for the novel objects (bottom). The last two examples may be considered to be failure cases (because the novel object classes, suitcase and bottle, are not mentioned).

captions. For consistency with previous work, out-of-domain scores are macro-averaged across the held-out classes, and CIDEr document frequency statistics are determined across the entire test set.

Results In Table 1 we show validation set results for the Up-Down model with various combinations of PS3 training and constrained beam search decoding (top panel), as well as performance upper bounds using ground-truth data (bottom panel). For constrained beam search decoding, image label predictions are generated by a linear mapping from the mean-pooled image feature 1

k Pk i=1 vi to image label scores which is trained on the entire training set. The results demonstrate that, on out-of-domain images, imposing the caption constraints during training using PS3 helps more than imposing the constraints during decoding. Furthermore, the model trained with PS3 has assimilated all the information available from the external image labeler, such that using constrained beam search during decoding provides no additional beneﬁt (row 3 vs. row 4). Overall, the model trained on image labels with PS3 (row 3) is closer in performance to the model trained with all captions (row 7) than it is to the baseline model (row 1). Evaluating our model (row 3) on the test set, we achieve state of the art results on the COCO novel object captioning task, as illustrated in Table 2. In Figure 3 we provide examples of generated captions, including failure cases. In Figure 4 we visualize attention in the model (suggesting that image label supervision can successfully train a visual attention mechanism to localize new objects).

5.2 Preliminary experiments on Open Images

Our primary motivation in this work is to extend the visual vocabulary of existing captioning models by making large object detection datasets available for training. Therefore, as a proof of concept

A woman holding a tennis racket on a court.

Figure 4: To further explore the impact of training using PS3, we visualize attention in the Up Down [10] model. As shown in this example, using only image label supervision (i.e., without caption supervision) the model still learns to ground novel object classes (such as racket) in the image.

tiger monkey rhino rabbit

Baseline: A zebra is laying down in the grass. Baseline: A black elephant laying on top of a wooden surface.

Baseline: A man taking a picture of an old car. Baseline: A cat that is laying on the grass.

Ours: A tiger that is sitting in the grass. Ours: A monkey that is sitting on the ground. Ours: A man sitting in a car looking at an elephant.

Ours: A squirrel that is sitting in the grass.

Figure 5: Preliminary experiments on Open Images. As expected, the baseline Up-Down [10] model trained on COCO performs poorly on novel object classes from the Open Images dataset (top). Incorporating image labels from 25 selected classes using PS3 leads to qualitative improvements (bottom). The last two examples are failure cases (but no worse than the baseline).

we train a captioning model simultaneously on COCO Captions [6] and object annotation labels for 25 additional animal classes from the Open Images V4 dataset [14]. In Figure 5 we provide some examples of the generated captions. We also evaluate the jointly trained model on the COCO Karpathy val split [27], achieving SPICE, METEOR and CIDEr scores of 18.8, 25.7 and 103.5, respectively, versus 20.1, 26.9 and 112.3 for the model trained exclusively on COCO.

6 Conclusion

We propose a novel algorithm for training sequence models on partially-speciﬁed data represented by ﬁnite state automata. Applying this approach to image captioning, we demonstrate that a generic image captioning model can learn new visual concepts from labeled images, achieving state of the art results on the COCO novel object captioning splits. We further show that we can train the model to describe new visual concepts from the Open Images dataset while maintaining competitive COCO evaluation scores. Future work could investigate training captioning models on ﬁnite state automata constructed from scene graph and visual relationship annotations, which are also available at large scale [14, 49].

Acknowledgments

This research was supported by a Google award through the Natural Language Understanding Focused Program, CRP 8201800363 from Data61/CSIRO, and under the Australian Research Council s Discovery Projects funding scheme (project number DP160102156). We also thank the anonymous reviewers for their valuable comments that helped to improve the paper.

[1] Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image caption generator. In CVPR, 2015.

[2] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron C. Courville, Ruslan Salakhutdinov, Richard S. Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In ICML, 2015.

[3] Hao Fang, Saurabh Gupta, Forrest N. Iandola, Rupesh Srivastava, Li Deng, Piotr Dollar, Jianfeng Gao, Xiaodong He, Margaret Mitchell, John C. Platt, C. Lawrence Zitnick, and Geoffrey Zweig. From captions to visual concepts and back. In CVPR, 2015.

[4] Micah Hodosh, Peter Young, and Julia Hockenmaier. Framing image description as a ranking task: Data, models and evaluation metrics. Journal of Artiﬁcial Intelligence Research, 47: 853 899, 2013.

[5] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67 78, 2014.

[6] Xinlei Chen, Tsung-Yi Lin Hao Fang, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C. Lawrence Zitnick. Microsoft COCO Captions: Data Collection and Evaluation Server. ar Xiv preprint ar Xiv:1504.00325, 2015.

[7] Steven J. Rennie, Etienne Marcheret, Youssef Mroueh, Jarret Ross, and Vaibhava Goel. Selfcritical sequence training for image captioning. In CVPR, 2017.

[8] Jiasen Lu, Caiming Xiong, Devi Parikh, and Richard Socher. Knowing when to look: Adaptive attention via a visual sentinel for image captioning. In CVPR, 2017.

[9] Zhilin Yang, Ye Yuan, Yuexin Wu, Ruslan Salakhutdinov, and William W. Cohen. Review networks for caption generation. In NIPS, 2016.

[10] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In CVPR, 2018.

[11] Kenneth Tran, Xiaodong He, Lei Zhang, Jian Sun, Cornelia Carapcea, Chris Thrasher, Chris Buehler, and Chris Sienkiewicz. Rich Image Captioning in the Wild. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2016.

[12] Haley Mac Leod, Cynthia L Bennett, Meredith Ringel Morris, and Edward Cutrell. Understanding blind people s experiences with computer-generated captions of social media images. In Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems. ACM, 2017.

[13] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Guided open vocabulary image captioning with constrained beam search. In EMNLP, 2017.

[14] Ivan Krasin, Tom Duerig, Neil Alldrin, Vittorio Ferrari, Sami Abu-El-Haija, Alina Kuznetsova, Hassan Rom, Jasper Uijlings, Stefan Popov, Andreas Veit, Serge Belongie, Victor Gomes, Abhinav Gupta, Chen Sun, Gal Chechik, David Cai, Zheyun Feng, Dhyanesh Narayanan, and Kevin Murphy. Openimages: A public dataset for large-scale multi-label and multi-class image classiﬁcation. Dataset available from https://github.com/openimages, 2017.

[15] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. International Journal of Computer Vision (IJCV), 2015.

[16] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. YFCC100M: the new data in multimedia research. Communications of the ACM, 59(2):64 73, 2016.

[17] Dim P Papadopoulos, Jasper RR Uijlings, Frank Keller, and Vittorio Ferrari. Extreme clicking for efﬁcient object annotation. In ICCV, 2017.

[18] Dim P Papadopoulos, Jasper RR Uijlings, Frank Keller, and Vittorio Ferrari. We don t need no bounding-boxes: Training object class detectors using only human veriﬁcation. In CVPR, 2016.

[19] Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statistical society. Series B (methodological), 1977.

[20] Geoffrey J. Mc Lachlan and Thriyambakam Krishnan. The EM algorithm and extensions. Wiley series in probability and statistics. Wiley, Hoboken, NJ, 2. ed edition, 2008. ISBN 978-0-471-20170-0.

[21] Lisa Anne Hendricks, Subhashini Venugopalan, Marcus Rohrbach, Raymond Mooney, Kate Saenko, and Trevor Darrell. Deep Compositional Captioning: Describing Novel Object Categories without Paired Training Data. In CVPR, 2016.

[22] Subhashini Venugopalan, Lisa Anne Hendricks, Marcus Rohrbach, Raymond J. Mooney, Trevor Darrell, and Kate Saenko. Captioning Images with Diverse Objects. In CVPR, 2017.

[23] Ting Yao, Yingwei Pan, Yehao Li, and Tao Mei. Incorporating copying mechanism in image captioning for learning novel objects. In CVPR, 2017.

[24] Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Neural baby talk. In CVPR, 2018.

[25] Jeffrey Donahue, Lisa A. Hendricks, Sergio Guadarrama, Marcus Rohrbach, Subhashini Venugopalan, Kate Saenko, and Trevor Darrell. Long-term recurrent convolutional networks for visual recognition and description. In CVPR, 2015.

[26] Junhua Mao, Wei Xu, Yi Yang, Jiang Wang, and Alan L. Yuille. Deep captioning with multimodal recurrent neural networks (m-rnn). In ICLR, 2015.

[27] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In CVPR, 2015.

[28] Jacob Devlin, Hao Cheng, Hao Fang, Saurabh Gupta, Li Deng, Xiaodong He, Geoffrey Zweig, and Margaret Mitchell. Language models for image captioning: The quirks and what works. ar Xiv preprint ar Xiv:1505.01809, 2015.

[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017.

[30] Alex Graves. Generating sequences with recurrent neural networks. ar Xiv preprint ar Xiv:1308.0850, 2013.

[31] Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. Sequence to sequence learning with neural networks. In NIPS, 2014.

[32] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. In ICLR, 2015.

[33] Ting Yao, Yingwei Pan, Yehao Li, and Tao Mei. Incorporating copying mechanism in image captioning for learning novel objects. In CVPR, 2017.

[34] Andrew M Dai and Quoc V Le. Semi-supervised sequence learning. In NIPS, 2015.

[35] Felix Hill, Kyunghyun Cho, and Anna Korhonen. Learning distributed representations of sentences from unlabelled data. ar Xiv preprint ar Xiv:1602.03483, 2016.

[36] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efﬁcient estimation of word representations in vector space. ar Xiv preprint ar Xiv:1301.3781, 2013.

[37] Jeffrey Pennington, Richard Socher, and Christopher D. Manning. Glo Ve: Global Vectors for Word Representation. In EMNLP, 2014.

[38] Shahla Parveen and Phil Green. Speech recognition with missing data using recurrent neural nets. In NIPS, 2002.

[39] Zachary C. Lipton, David C. Kale, and Randall Wetzel. Modeling missing data in clinical time series with RNNs. In Machine Learning for Healthcare, 2016.

[40] Zoubin Ghahramani and Michael I Jordan. Supervised learning from incomplete data via an em approach. In NIPS, 1994.

[41] Michael Sipser. Introduction to the Theory of Computation. Cengage Learning, 3rd edition, 2012.

[42] Philipp Koehn. Statistical Machine Translation. Cambridge University Press, New York, NY, USA, 1st edition, 2010. ISBN 0521874157, 9780521874151.

[43] Chris Hokamp and Qun Liu. Lexically constrained decoding for sequence generation using grid beam search. In ACL, 2017.

[44] Matt Post and David Vilar. Fast lexically constrained decoding with dynamic beam allocation for neural machine translation. ar Xiv preprint ar Xiv:1804.06609, 2018.

[45] Kyle Richardson, Jonathan Berant, and Jonas Kuhn. Polyglot semantic parsing in APIs. In NAACL, 2018.

[46] Sam Wiseman and Alexander M Rush. Sequence-to-sequence learning as beam-search optimization. In EMNLP, 2016.

[47] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NIPS, 2015.

[48] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016.

[49] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, Michael Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. ar Xiv preprint ar Xiv:1602.07332, 2016.

[50] Sepp Hochreiter and Jürgen Schmidhuber. Long Short-Term Memory. Neural Computation, 1997.

[51] Omer Levy and Yoav Goldberg. Dependency-based word embeddings. In ACL, 2014.

[52] T.Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollar, and C. L. Zitnick. Microsoft COCO: Common Objects in Context. In ECCV, 2014.

[53] Yongqin Xian, Bernt Schiele, and Zeynep Akata. Zero-shot learning-the good, the bad and the ugly. In CVPR, 2017.

[54] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. SPICE: Semantic Propositional Image Caption Evaluation. In ECCV, 2016.

[55] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. CIDEr: Consensus-based image description evaluation. In CVPR, 2015.

[56] Alon Lavie and Abhaya Agarwal. Meteor: An automatic metric for MT evaluation with high levels of correlation with human judgments. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL): Second Workshop on Statistical Machine Translation, 2007.