Sequence to Sequence Video to Text Subhashini Venugopalan, Marcus Rohrbach, Jeff Donahue, Raymond Mooney, Trevor Darrell, Kate Saenko Presented by Dewal Gupta UCSD CSE 291G, Winter 2019
BACKGROUND Challenge: Create a description for a given video Important in: describing videos for blind human-robot interactions Challenging because: diverse set of scenes, actions necessary to recognize salient action in context images don’t have to face the
PREVIOUS WORK: Template Models Tag video with captions and use as bag of words Two stage pipeline: first: tag video with semantic information on objects, actions treated as a classification problem FGM labels subject, verb, object, place second: generate sentence from semantic information S2VT approach: avoids separating content identification from sentence generation Integrating Language and Vision: to Generate Natural Language Descriptions of Videos in the Wild - Mooney et al., 2014
PREVIOUS WORK: Mean Pooling CNN trained on object classification (subset of ImageNet) 2 layer LSTM with video and previous word as input Ignores video frame ordering Translating Videos to Natural Language Using Deep Recurrent Neural Networks Mooney et al., 2015
PREVIOUS WORK: Exploiting Temporal Structure Encoder: train 3D ConvNet on action recognition fixed frame input exploits local temporal structure Describing Videos by Exploiting Temporal Structure Videos in the Wild Courville et al., 2015
PREVIOUS WORK: Exploiting Temporal Structure Decoder: Similar to our HW 2 Exploits global temporal structure Describing Videos by Exploiting Temporal Structure Videos in the Wild Courville et al., 2015
GOAL End to End differentiable model that can: Handle variable video length (i.e. variable input length) Learn temporal structure Learn a language model that is capable of generating descriptive sentences
MODEL: LSTM Single LSTM network 2 layer LSTM network 1000 hidden units (ht) red layer: models visual elements green layer: models linguistic elements ILSVRC-2012 object classification subset of the ImageNet dataset [30]
MODEL: VGG-16
MODEL: AlexNet Used for RGB & Flow!
MODEL: Details Use Text Embedding (of 500 dimensions) self-trained, simple linear transformation RGB networks are pre-trained on subset of ImageNet data Used networks from the original works Optical Flow pretrained on UCF101 dataset Action Classification Task Original work from ‘Action Tubes’ All layers are frozen except last layers for training Flow and RGB combined by “shallow fusion technique” alpha is tuned on the validation set
DATASETS 3 datasets used: Microsoft Video Description corpus (MSVD) MPII Movie Description Corpus (MPII-MD) Montreal Video Annotation Dataset (M-VAD) MSVD: web clips with human annotations MPII-MD: Hollywood clips with descriptions from script & audio (originally for the visually impaired) M-VAD: Hollywood clips with audio descriptions All three have single sentence descriptions
DATASETS: Metrics Authors use METEOR metric uses exact token, stemmed token and WordNet synonym matches better correlation with human judgement than BLEU or ROUGE out performs CIDEr when fewer references datasets only had 1 reference where: m is unigram (or n-gram) matches after alignment wr is length of reference wt is length of candidate BLEU does not take recall into consideration directly - rather just penalizes brevity
RESULTS: MSVD FGM is template based not very descriptive predicts a noun, verb, object, place builds sentence off template
RESULTS: MSVD Mean Pool based method very similar to author’s method
RESULTS: MSVD Temporal Attention method Encoder/Decoder using attention
RESULTS: Frame ordering Training with random ordering of frames results in “considerably lower” performance
RESULTS: Optical Flow Flow results in better performance only when combined with RGB (& not when used alone) Flow can be very different even for same activities Flow can’t account for polysemous words like “play” - eg. “play guitar” vs “play golf” person vs panda eating
RESULTS: SOTA Authors claim accurate comparison is with GoogleNet with NO 3D-CNN (global temporal attention) questionable claim person vs panda eating
Results: MPII-MD, M-VAD Similar performance to Visual-Labels VL uses more semantic information (eg. object detection) but no temporal information
Results: Edit Distance Levenshtein Distance: represents edit distance between two strings 42.9% of generated samples match exactly with a sentence in the training corpus of MSVD model struggles to learn MVAD
CRITICISM Model fails to learn temporal relations performs nearly as well as mean pooling technique that makes no use of temporal relations Model struggles on MVAD dataset for some reason more than other Authors should have used BLEU and/or CIDEr scores as well (other studies have them) Conduct user study (where human looks at captions)? Could improve by using better text embeddings?
FURTHER WORK Use Inception ResNet v2 as backbone CNN Train CNN against mined video “attributes” Achieve +5% METEOR score on MSVD Same architecture End-to-End Video Captioning with Multitask Reinforcement Learning - Li & Gong, 2019
FURTHER WORK Use 3D CNN to get better clip embeddings instead of LSTMs proven better in activity recognition Rethinking Spatiotemporal Feature Learning: Speed-Accuracy Trade-offs in Video Classification - Xie et al., 2017
CONCLUSION Authors build an end to end differentiable model that can: Handle variable video length (i.e. variable input length) Learn temporal structure Learn a language model that is capable of generating descriptive sentences Has become a baseline for many video captioning works
EXAMPLES