Research map / Detection and segmentation / CNN features
Predicting Future Instance Segmentation by Forecasting Convolutional Features
Pauline Luc, Camille Couprie, Yann LeCun, Jakob Verbeek
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 |
|---|---|---|---|
| add P4 features (use P4-P5 instead of P5 only) (ablation) | F2F · Cityscapes val, short-term instance segmentation | AP50: 2.2 -> 10.2 | Improved |
| add P3 features (use P3-P5 instead of P4-P5) (ablation) | F2F · Cityscapes val, short-term instance segmentation | AP50: 10.2 -> 24.8 | Improved |
| add P2 features (use P2-P5 instead of P3-P5) (ablation) | F2F · Cityscapes val, short-term instance segmentation | AP50: 24.8 -> 40.2 | Improved |
| train a specific network per feature level instead of sharing the P5 network across levels (ablation) | F2F · Cityscapes val, short-term instance segmentation | AP50: 16.7 -> 40.2 | Improved |
| add autoregressive fine-tuning (ablation) | F2F · Cityscapes val, mid-term instance segmentation | AP50: 17.5 -> 19.4 | Improved |
| add autoregressive fine-tuning (ablation) | F2F · Cityscapes val, mid-term instance segmentation | AP: 6.2 -> 7.7 | Improved |
| add autoregressive fine-tuning (ablation) | F2F · Cityscapes val, short-term instance segmentation | AP50: 40.2 -> 39.9 | No clear effect |
| predict future in feature space instead of RGB space then segment | F2F · Cityscapes val, short-term instance segmentation | AP50: 6.9 -> 40.2 | Improved |
| fine-tune Mask R-CNN on blurry predicted RGB frames (ablation) | RGB-prediction baseline · Cityscapes val, short-term instance segmentation | AP50: 6.9 -> 19.2 | Improved |
| add morphological closing post-processing to Warp (ablation) | Warp · Cityscapes val, mid-term instance segmentation | AP50: 5.7 -> 10.9 | Improved |
| add morphological opening to Warp post-processing (ablation) | Warp · Cityscapes val, mid-term instance segmentation | AP50: 10.9 -> 11.1 | Improved |
| use flow warping instead of mean-shift of masks | Optical flow baseline · Cityscapes val, mid-term semantic segmentation | IoU: 36.7 -> 41.4 | Improved |
Closest papers by meaning
- Predicting Deeper Into the Future of Semantic Segmentation (ICCV 2017 · similarity 0.92)
- SOLO: Segmenting Objects by Locations (ECCV 2020 · similarity 0.88)
- APANet: Auto-Path Aggregation for Future Instance Segmentation Prediction (TPAMI 2021 · similarity 0.88)
- Probabilistic Future Prediction for Video Scene Understanding (ECCV 2020 · similarity 0.87)
- SOLO: A Simple Framework for Instance Segmentation (TPAMI 2021 · similarity 0.86)
- Video Instance Segmentation (ICCV 2019 · similarity 0.86)
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.
