SwinTransformer / Video-Swin-Transformer

This is an official implementation for "Video Swin Transformers".
https://arxiv.org/abs/2106.13230
Apache License 2.0
1.44k stars 199 forks source link

Question about Table 5. Ablation study on the temporal dimension of 3D tokens and temporal window size with Swin-T on K400 and SSv2 #95

Open pbhfcycssjlmm opened 4 months ago

pbhfcycssjlmm commented 4 months ago

What is the difference between temporal dimension of 3D tokens and temporal window size?

Or

"Fixing the temporal dimension of 3D tokens to 16, we perform an ablation study over temporal window sizes of 4/8/16." How is the temporal dimension fixed to 16 to support setting different temporal window sizes?

pbhfcycssjlmm commented 4 months ago

Does temporal window size refer to 3D shifted windows size? Corresponding to PMM mentioned in Figure 3?

And temporal dimension of 3D tokens refers to the time step, corresponding to the 2 mentioned in "Unless otherwise mentioned, for all model variants, we sample a clip of 32 frames from each full length video using a temporal stride of 2 and spatial size of 224 × 224, resulting in 16×56×56 input 3D tokens."? But why are 16, 8, and 4 used in the experiment?