PatchBlender: A Motion Prior for Video Transformers
This addresses a specific bottleneck in video analysis for computer vision researchers, offering an incremental improvement with lightweight computation.
The authors tackled the challenge of vision transformers struggling to model temporal patterns in video data by introducing PatchBlender, a learnable blending function for patch embeddings, which improved baseline performance on Something-Something v2 and MOVi-A datasets.
Transformers have become one of the dominant architectures in the field of computer vision. However, there are yet several challenges when applying such architectures to video data. Most notably, these models struggle to model the temporal patterns of video data effectively. Directly targeting this issue, we introduce PatchBlender, a learnable blending function that operates over patch embeddings across the temporal dimension of the latent space. We show that our method is successful at enabling vision transformers to encode the temporal component of video data. On Something-Something v2 and MOVi-A, we show that our method improves the baseline performance of video Transformers. PatchBlender has the advantage of being compatible with almost any Transformer architecture and since it is learnable, the model can adaptively turn on or off the prior. It is also extremely lightweight compute-wise, 0.005% the GFLOPs of a ViT-B.