Research map / Video / Action recognition
FrameExit: Conditional Early Exiting for Efficient Video Recognition
Check 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 |
|---|---|---|---|
| use conditional early exiting instead of processing all frames (ablation) | FrameExit with ResNet-50 · ActivityNet-v1.3 | GFLOPs: 41.2 -> 26.1 | Helped |
| use conditional early exiting instead of processing all frames (ablation) | FrameExit with ResNet-50 · ActivityNet-v1.3 | mAP (%): 77.3 -> 76.1 | Hurt |
| use deterministic coarse-to-fine frame sampling instead of sequential sampling (ablation) | FrameExit · ActivityNet-v1.3 | mAP (%): 73.7 -> 76.1 | Helped |
| randomize sampling during training and testing instead of sequential sampling (ablation) | FrameExit · ActivityNet-v1.3 | mAP (%): 73.7 -> 74.4 ± 0.34 | Helped |
| add accumulated feature pooling rather than pooling frame-level predictions (ablation) | FrameExit · ActivityNet-v1.3 | mAP (%): 67.8 -> 76.1 | Helped |
| increase the maximum input frames from 10 to 16 (ablation) | FrameExit · ActivityNet-v1.3 | mAP (%): 76.1 -> 76.1 | No clear effect |
| increase the maximum input frames from 10 to 16 (ablation) | FrameExit · ActivityNet-v1.3 | GFLOPs: 26.1 -> 35.1 | Hurt |
| replace adaptive exiting with a fixed 3-frame budget (ablation) | FrameExit · ActivityNet-v1.3 | accuracy: 60.8 -> 67.5 | Hurt |
| replace adaptive exiting with a fixed 4-frame budget (ablation) | FrameExit · ActivityNet-v1.3 | accuracy: 67.5 -> 70.1 | Hurt |
| use FrameExit with EfficientNet-b3 instead of AR-Net | EfficientNet-b3 · ActivityNet-v1.3 | GFLOPs: 15.3 -> 11.4 | Helped |
| use FrameExit with EfficientNet-b3 instead of AR-Net | EfficientNet-b3 · Mini-Kinetics | GFLOPs: 16.3 -> 7.8 | Helped |
| raise β from 1e^{-3} to 1e^{-5} | FrameExit with ResNet-50 · HVU | mAP (%): 45.7 -> 49.2 | Helped |
Closest papers by meaning
- AdaFrame: Adaptive Frame Selection for Fast Video Recognition (CVPR 2019 · similarity 0.88)
- LookWhen? Fast Video Recognition by Learning When, Where, and What to Compute (NeurIPS 2026 · similarity 0.87)
- SMART Frame Selection for Action Recognition (AAAI 2021 · similarity 0.87)
- Smooth Regularization for Efficient Video Recognition (NeurIPS 2025 · similarity 0.85)
- No More Shortcuts: Realizing the Potential of Temporal Self-Supervision (AAAI 2024 · similarity 0.84)
- TTVidT: Decoupling the Temporal Axis for Efficient Motion-Centric Video Pretraining (NeurIPS 2026 · similarity 0.84)
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.