Keeping Your Eye on the Ball: Trajectory Attention in Video Transformers

In video transformers, the time dimension is often treated in the same way as the two spatial dimensions. However, in a scene where objects or the camera may move, a physical point imaged at one location in frame $t$ may be entirely unrelated to what is found at that location in frame $t+k$... These temporal correspondences should be modeled to facilitate learning about dynamic scenes. To this end, we propose a new drop-in block for video transformers -- trajectory attention -- that aggregates information along implicitly determined motion paths. We additionally propose a new method to address the quadratic dependence of computation and memory on the input size, which is particularly important for high resolution or long videos. While these ideas are useful in a range of settings, we apply them to the specific task of video action recognition with a transformer model and obtain state-of-the-art results on the Kinetics, Something--Something V2, and Epic-Kitchens datasets. Code and models are available at: https://github.com/facebookresearch/Motionformer read more

PDF Abstract NeurIPS 2021 PDF NeurIPS 2021 Abstract

Results from the Paper Edit

Ranked #4 on Action Recognition on EPIC-KITCHENS-100 (using extra training data)

Task Dataset Model Metric Name Metric Value Global Rank Uses Extra
Training Data
Result Benchmark
Action Recognition EPIC-KITCHENS-100 Mformer-L Action@1 44.1 # 7
Verb@1 67.1 # 5
Noun@1 57.6 # 4
Action Recognition EPIC-KITCHENS-100 Mformer Action@1 43.1 # 10
Verb@1 66.7 # 8
Noun@1 56.5 # 7
Action Recognition EPIC-KITCHENS-100 Mformer-HR Action@1 44.5 # 4
Verb@1 67.0 # 7
Noun@1 58.5 # 2
Action Classification Kinetics-400 Motionformer-HR Vid acc@1 81.1 # 18
Vid acc@5 95.2 # 9
Action Recognition Something-Something V2 Mformer Top-1 Accuracy 66.5 # 19
Top-5 Accuracy 90.1 # 22
Action Recognition Something-Something V2 Mformer-L(IN-21K+Kinetics400 pretrain) Top-1 Accuracy 68.1 # 6
Top-5 Accuracy 91.2 # 10
Parameters N/A # 14
GFLOPs 1181x3 # 1
Action Recognition Something-Something V2 Mformer-HR(IN-21K+Kinetics400 pretrain) Top-1 Accuracy 67.1 # 15
Top-5 Accuracy 90.6 # 14
Parameters N/A # 14
GFLOPs 958.8x3 # 1