This paper introduces EPIC-KITCHENS-100, the largest annotated egocentric dataset - 100 hrs, 20M frames, 90K actions - of wearable videos capturing long-term unscripted activities in 45 environments. This extends our previous dataset (EPIC-KITCHENS-55), released in 2018, resulting in more action segments (+128%), environments (+41%) and hours (+84%), using a novel annotation pipeline that allows denser and more complete annotations of fine-grained actions (54% more actions per minute). We evaluate the "test of time" - i.e. whether models trained on data collected in 2018 can generalise to new footage collected under the same hypotheses albeit "two years on". The dataset is aligned with 6 challenges: action recognition (full and weak supervision), detection, anticipation, retrieval (from captions), as well as unsupervised domain adaptation for action recognition. For each challenge, we define the task, provide baselines and evaluation metrics. Our dataset and challenge leaderboards will be made publicly available.
Since its introduction in 2018, EPIC-KITCHENS has attracted attention as the largest egocentric video benchmark, offering a unique viewpoint on people's interaction with objects, their attention, and even intention. In this paper, we detail how this large-scale dataset was captured by 32 participants in their native kitchen environments, and densely annotated with actions and object interactions. Our videos depict nonscripted daily activities, as recording is started every time a participant entered their kitchen. Recording took place in 4 countries by participants belonging to 10 different nationalities, resulting in highly diverse kitchen habits and cooking styles. Our dataset features 55 hours of video consisting of 11.5M frames, which we densely labelled for a total of 39.6K action segments and 454.2K object bounding boxes. Our annotation is unique in that we had the participants narrate their own videos after recording, thus reflecting true intention, and we crowd-sourced ground-truths based on these. We describe our object, action and. anticipation challenges, and evaluate several baselines over two test splits, seen and unseen kitchens. We introduce new baselines that highlight the multimodal nature of the dataset and the importance of explicit temporal modelling to discriminate fine-grained actions e.g. 'closing a tap' from 'opening' it up.
Fine-grained action recognition datasets exhibit environmental bias, where multiple video sequences are captured from a limited number of environments. Training a model in one environment and deploying in another results in a drop in performance due to an unavoidable domain shift. Unsupervised Domain Adaptation (UDA) approaches have frequently utilised adversarial training between the source and target domains. However, these approaches have not explored the multi-modal nature of video within each domain. In this work we exploit the correspondence of modalities as a self-supervised alignment approach for UDA in addition to adversarial alignment. We test our approach on three kitchens from our large-scale dataset, EPIC-Kitchens, using two modalities commonly employed for action recognition: RGB and Optical Flow. We show that multi-modal self-supervision alone improves the performance over source-only training by 2.4% on average. We then combine adversarial training with multi-modal self-supervision, showing that our approach outperforms other UDA methods by 3%.
We present a method to learn a representation for adverbs from instructional videos using weak supervision from the accompanying narrations. Key to our method is the fact that the visual representation of the adverb is highly dependant on the action to which it applies, although the same adverb will modify multiple actions in a similar way. For instance, while 'spread quickly' and 'mix quickly' will look dissimilar, we can learn a common representation that allows us to recognize both, among other actions. We formulate this as an embedding problem, and use scaled dot-product attention to learn from weakly-supervised video narrations. We jointly learn adverbs as invertible transformations operating on the embedding space, so as to add or remove the effect of the adverb. As there is no prior work on weakly supervised learning from adverbs, we gather paired action-adverb annotations from a subset of the HowTo100M dataset for 6 adverbs: quickly/slowly, finely/coarsely, and partially/completely. Our method outperforms all baselines for video-to-adverb retrieval with a performance of 0.719 mAP. We also demonstrate our model's ability to attend to the relevant video parts in order to determine the adverb for a given action.
Monitoring the progression of an action towards completion offers fine grained insight into the actor's behaviour. In this work, we target detecting the completion moment of actions, that is the moment when the action's goal has been successfully accomplished. This has potential applications from surveillance to assistive living and human-robot interactions. Previous effort required human annotations of the completion moment for training (i.e. full supervision). In this work, we present an approach for moment detection from weak video-level labels. Given both complete and incomplete sequences, of the same action, we learn temporal attention, along with accumulated completion prediction from all frames in the sequence. We also demonstrate how the approach can be used when completion moment supervision is available. We evaluate and compare our approach on actions from three datasets, namely HMDB, UCF101 and RGBD-AC, and show that temporal attention improves detection in both weakly-supervised and fully-supervised settings.
We present the first fully automated Sit-to-Stand or Stand-to-Sit (StS) analysis framework for long-term monitoring of patients in free-living environments using video silhouettes. Our method adopts a coarse-to-fine time localisation approach, where a deep learning classifier identifies possible StS sequences from silhouettes, and a smart peak detection stage provides fine localisation based on 3D bounding boxes. We tested our method on data from real homes of participants and monitored patients undergoing total hip or knee replacement. Our results show 94.4% overall accuracy in the coarse localisation and an error of 0.026 m/s in the speed of ascent measurement, highlighting important trends in the recuperation of patients who underwent surgery.
We investigate video transforms that result in class-homogeneous label-transforms. These are video transforms that consistently maintain or modify the labels of all videos in each class. We propose a general approach to discover invariant classes, whose transformed examples maintain their label; pairs of equivariant classes, whose transformed examples exchange their labels; and novel-generating classes, whose transformed examples belong to a new class outside the dataset. Label transforms offer additional supervision previously unexplored in video recognition benefiting data augmentation and enabling zero-shot learning opportunities by learning a class from transformed videos of its counterpart. Amongst such video transforms, we study horizontal-flipping, time-reversal, and their composition. We highlight errors in naively using horizontal-flipping as a form of data augmentation in video. Next, we validate the realism of time-reversed videos through a human perception study where people exhibit equal preference for forward and time-reversed videos. Finally, we test our approach on two datasets, Jester and Something-Something, evaluating the three video transforms for zero-shot learning and data augmentation. Our results show that gestures such as zooming in can be learnt from zooming out in a zero-shot setting, as well as more complex actions with state transitions such as digging something out of something from burying something in something.
We focus on multi-modal fusion for egocentric action recognition, and propose a novel architecture for multi-modal temporal-binding, i.e. the combination of modalities within a range of temporal offsets. We train the architecture with three modalities -- RGB, Flow and Audio -- and combine them with mid-level fusion alongside sparse temporal sampling of fused representations. In contrast with previous works, modalities are fused before temporal aggregation, with shared modality and fusion weights over time. Our proposed architecture is trained end-to-end, outperforming individual modalities as well as late-fusion of modalities. We demonstrate the importance of audio in egocentric vision, on per-class basis, for identifying actions as well as interacting objects. Our method achieves state of the art results on both the seen and unseen test sets of the largest egocentric dataset: EPIC-Kitchens, on all metrics using the public leaderboard.
We address the problem of cross-modal fine-grained action retrieval between text and video. Cross-modal retrieval is commonly achieved through learning a shared embedding space, that can indifferently embed modalities. In this paper, we propose to enrich the embedding by disentangling parts-of-speech (PoS) in the accompanying captions. We build a separate multi-modal embedding space for each PoS tag. The outputs of multiple PoS embeddings are then used as input to an integrated multi-modal space, where we perform action retrieval. All embeddings are trained jointly through a combination of PoS-aware and PoS-agnostic losses. Our proposal enables learning specialised embedding spaces that offer multiple views of the same embedded entities. We report the first retrieval results on fine-grained actions for the large-scale EPIC dataset, in a generalised zero-shot setting. Results show the advantage of our approach for both video-to-text and text-to-video action retrieval. We also demonstrate the benefit of disentangling the PoS for the generic task of cross-modal video retrieval on the MSR-VTT dataset.
We benchmark contemporary action recognition models (TSN, TRN, and TSM) on the recently introduced EPIC-Kitchens dataset and release pretrained models on GitHub (https://github.com/epic-kitchens/action-models) for others to build upon. In contrast to popular action recognition datasets like Kinetics, Something-Something, UCF101, and HMDB51, EPIC-Kitchens is shot from an egocentric perspective and captures daily actions in-situ. In this report, we aim to understand how well these models can tackle the challenges present in this dataset, such as its long tail class distribution, unseen environment test set, and multiple tasks (verb, noun and, action classification). We discuss the models' shortcomings and avenues for future research.