Research map / Detection and segmentation / 3D detection and driving
A Dynamic Feature Interaction Framework for Multi-task Visual Perception
Yuling Xi, Hao Chen, Ning Wang, Peng Wang, Yanning Zhang, Chunhua Shen, Yifan Liu
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 Dynamic Message Passing (DMP) to the two-branch baseline (ablation) | D2BNet · Cityscapes val multi-task | DS: 45.38 -> 45.73 | Improved |
| add task-aware dynamic router (TDR) to DMP (ablation) | D2BNet · Cityscapes val multi-task | DS: 45.73 -> 47.09 | Improved |
| add channel-aware dynamic router (CDR) to DMP (ablation) | D2BNet · Cityscapes val multi-task | DS: 45.73 -> 47.6 | Improved |
| add CDR on top of DMP and TDR (ablation) | D2BNet · Cityscapes val multi-task | PQ: 58.13 -> 58.6 | Improved |
| add DMP for single-task panoptic segmentation (ablation) | D2BNet · Cityscapes panoptic (single task) | PQ: 58.7 -> 60.0 | Improved |
| co-train panoptic segmentation with depth (no dynamic modules) (ablation) | D2BNet · Cityscapes val | PQ: 58.23 -> 56.40 | Hurt |
| co-train depth with segmentation (no dynamic modules) (ablation) | D2BNet · Cityscapes val | Abs. Rel.: 0.124 -> 0.111 | Improved |
| alternate training over nuScenes and nuImages | D2BNet · nuScenes | 3D mAP: 28.79 -> 30.63 | Improved |
| alternate training over nuScenes and nuImages | D2BNet · nuImages | mask mAP: 42.28 -> 39.81 | Hurt |
| compute context information from box tower instead of no context (ablation) | D2BNet · COCO instance segmentation (ablation) | AP: 34.7 -> 35.2 | Improved |
| use factored attention instead of channel-wise vector attention (ablation) | D2BNet · COCO instance segmentation (ablation) | AP: 35.2 -> 36.3 | Improved |
| double dense branch width from 32 to 64 (ablation) | D2BNet · COCO panoptic (R-50, 3x) | PQ: 41.8 -> 42.9 | Improved |
Closest papers by meaning
- RepVF: A Unified Vector Fields Representation for Multi-task 3D Perception (ECCV 2024 · similarity 0.89)
- Efficient Computation Sharing for Multi-Task Visual Scene Understanding (ICCV 2023 · similarity 0.88)
- TaskExpert: Dynamically Assembling Multi-Task Representations with Memorial Mixture-of-Experts (ICCV 2023 · similarity 0.87)
- BridgeNet: Comprehensive and Effective Feature Interactions via Bridge Feature for Multi-Task Dense Predictions (TPAMI 2025 · similarity 0.87)
- Taskology: Utilizing Task Relations at Scale (CVPR 2021 · similarity 0.87)
- Swiss Army Knife: Synergizing Biases in Knowledge from Vision Foundation Models for Multi-Task Learning (ICLR 2025 · 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.
