Research map / Label-efficient and robust learning / Self-supervised and contrastive
Time Does Tell: Self-Supervised Time-Tuning of Dense Image Representations
Mohammadreza Salehi, Efstratios Gavves, Cees G.M. Snoek, Yuki M. Asano
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| train on multiple frames with feature forwarding instead of single frames (ablation) | TimeT · Pascal VOC K=500 | mIoU: 26.2 -> 42.8 | Improved |
| forward Sinkhorn-Knopp clustered features instead of raw logits (ablation) | Feature-Forwarder · Pascal VOC K=500 | mIoU: 35.5 -> 42.8 | Improved |
| forward SK-clustered features instead of identity (static) forwarding (ablation) | Feature-Forwarder · Pascal VOC K=500 | mIoU: 36.1 -> 42.8 | Improved |
| set time interval to 0.5s instead of 0.2s or 1.0s (ablation) | TimeT · Pascal VOC K=500 | mIoU: 36.5 / 38.1 -> 42.8 | Improved |
| increase number of frames from 4 to 8 (ablation) | TimeT · Pascal VOC K=500 | mIoU: 42.8 -> 42.3 | No clear effect |
| increase prototypes from 10 to 200 (ablation) | TimeT · Pascal VOC K=500 | mIoU: 24.1 -> 42.8 | Improved |
| increase prototypes from 200 to 300 (ablation) | TimeT · Pascal VOC K=500 | mIoU: 42.8 -> 42.3 | No clear effect |
| apply TimeT to different pretrained backbones (MSN, iBOT, DINO ViT-S/8, DINO ViT-B/16) | TimeT · Pascal VOC K=500 | mIoU: 26.0 -> 48.3 (MSN) | Improved |
| use feature forwarder instead of optical flow for mask propagation (ablation) | Feature-Forwarder · DAVIS validation | mIoU: 28.5 -> 40.3 | Improved |
| fine-tune image-based Leopart on video frames instead of images | Leopart · Pascal VOC linear classification | mIoU: 68.0 -> 53.2 | Hurt |
| time-tuning DINO with video clustering loss versus DINO alone on video segmentation | DINO · DAVIS clustering per-frame | mIoU: 30.2 -> 53.7 | Improved |
| TimeT trained on video versus image-trained Leopart on Pascal K=21 | TimeT · Pascal VOC K=21 | mIoU: 36.6 -> 34.5 | No clear effect |
Closest papers by meaning
- TimeBridge: Self-Supervised Video Representation Learning via Start-End Joint Embedding and In-Between Frame Prediction (CVPR 2026 · similarity 0.90)
- Learning by Aligning Videos in Time (CVPR 2021 · similarity 0.89)
- Domain Adaptive Video Segmentation via Temporal Pseudo Supervision (ECCV 2022 · similarity 0.89)
- No More Shortcuts: Realizing the Potential of Temporal Self-Supervision (AAAI 2024 · similarity 0.88)
- Rethinking Self-Supervised Semantic Segmentation: Achieving End-to-End Segmentation (TPAMI 2024 · similarity 0.88)
- UniViT: Unifying Image and Video Understanding in One Vision Encoder (NeurIPS 2025 · 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.
