Composable Augmentation Encoding for Video Representation Learning

• Published in: 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 8814 - 8824
• Main Authors: Sun, Chen, Nagrani, Arsha, Tian, Yonglong, Schmid, Cordelia
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.10.2021
• Subjects: Benchmark testing; Computer vision; Data models; Encoding; Recognition and classification; Representation learning; Task analysis; Transfer/Low-shot/Semi/Unsupervised Learning; Transformers; Video analysis and understanding
• ISSN: 2380-7504
• DOI: 10.1109/ICCV48922.2021.00871

We focus on contrastive methods for self-supervised video representation learning. A common paradigm in contrastive learning is to construct positive pairs by sampling different data views for the same instance, with different data instances as negatives. These methods implicitly assume a set of representational invariances to the view selection mechanism (e.g., sampling frames with temporal shifts), which may lead to poor performance on downstream tasks which violate these invariances (fine-grained video action recognition that would benefit from temporal information). To overcome this limitation, we propose an 'augmentation aware' contrastive learning framework, where we explicitly provide a sequence of augmentation parameterisations (such as the values of the time shifts used to create data views) as composable augmentation encodings (CATE) to our model when projecting the video representations for contrastive learning. We show that representations learned by our method encode valuable information about specified spatial or temporal augmentation, and in doing so also achieve state-of-the-art performance on a number of video benchmarks.