Research map / Video / Action recognition
Trajectory Convolution for Action Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add trajectory convolution (avg-pooled TV-L1 flow) at res3b1.conv1 (ablation) | Separable-3D ResNet-18 · Something-Something V1 val | Top-1 accuracy: 34.30 -> 34.96 | Helped |
| use max pooling instead of average pooling to downsample flow (ablation) | TrajectoryNet at res3b1.conv1 · Something-Something V1 val | Top-1 accuracy: 34.96 -> 36.04 | Helped |
| add a second trajectory convolution (res3b1.conv1,2) (ablation) | TrajectoryNet · Something-Something V1 val | Top-1 accuracy: 34.96 -> 34.72 top-1; 66.24 -> 66.89 top-5 | Mixed |
| concatenate 4-dim trajectory coordinates with appearance features (ablation) | TrajectoryNet · Something-Something V1 val | Top-1 accuracy: 34.96 -> 36.08 | Helped |
| concatenate motion feature map instead of trajectory coordinates (ablation) | TrajectoryNet · Something-Something V1 val | Top-1 accuracy: 34.96 -> 35.24 | Helped |
| replace TV-L1 trajectories with frozen MotionNet-(2) (ablation) | TrajectoryNet · Something-Something V1 val | Top-1 accuracy: 34.96 -> 36.37 | Helped |
| fine-tune MotionNet-(2) with classification loss only (ablation) | TrajectoryNet · Something-Something V1 val | Top-1 accuracy: 36.37 -> 34.72 | Hurt |
| add unsupervised loss when fine-tuning MotionNet-(2) (ablation) | TrajectoryNet · Something-Something V1 val | Top-1 accuracy: 34.72 -> 36.91 | Helped |
| add unsupervised loss when fine-tuning MotionNet-(17) (ablation) | TrajectoryNet · Something-Something V1 val | Top-1 accuracy: 35.25 -> 36.69 | Helped |
| fine-tune MotionNet-(17) with classification loss only vs frozen (ablation) | TrajectoryNet · Something-Something V1 val | Top-1 accuracy: 35.69 -> 35.25 | Hurt |
| add trajectory convolution with MotionNet-(17) at sampling step 2 (ablation) | S3D 16 x 2 · Something-Something V1 val | Top-1 accuracy: 42.47 -> 43.32 | Helped |
| add trajectory convolution with MotionNet-(2) on Kinetics | Separable-3D ResNet-18 · Kinetics val | Average of top-1 and top-5 accuracy: 76.9 -> 77.8 | Helped |
Closest papers by meaning
- T-C3D: Temporal Convolutional 3D Network for Real-Time Action Recognition (AAAI 2018 · similarity 0.89)
- Temporal Convolutional Networks for Action Segmentation and Detection (CVPR 2017 · similarity 0.89)
- Deep Analysis of CNN-Based Spatio-Temporal Representations for Action Recognition (CVPR 2021 · similarity 0.88)
- ACTION-Net: Multipath Excitation for Action Recognition (CVPR 2021 · similarity 0.88)
- Memory-Augmented Temporal Dynamic Learning for Action Recognition (AAAI 2019 · similarity 0.88)
- MS-TCN: Multi-Stage Temporal Convolutional Network for Action Segmentation (CVPR 2019 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.