Research map / Detection and segmentation / CNN features
Revisiting Multi-Task Learning with ROCK: a Deep Residual Auxiliary Block for Visual Detection
Taylor Mordan, Nicolas THOME, Gilles Henaff, Matthieu Cord
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 residual auxiliary block (ROCK) with auxiliary supervision to SSD | SSD ResNet-50 · NYUv2 val | mAP@0.5: 31.2 -> 37.6 | Improved |
| add auxiliary annotations via flat MTL heads (ablation) | SSD ResNet-50 · NYUv2 val | mAP@0.5: 31.2 -> 34.3 | Improved |
| add ROCK auxiliary block encoder/predictor without fusion over flat MTL (ablation) | flat MTL SSD · NYUv2 val | mAP@0.5: 34.3 -> 35.7 | Improved |
| add feature merging (decoder and fusion) (ablation) | ROCK w/o fusion · NYUv2 val | mAP@0.5: 35.7 -> 37.6 | Improved |
| add scene classification supervision (ablation) | ROCK · NYUv2 val | mAP@0.5: 34.0 -> 37.6 | Improved |
| add normal supervision (ablation) | ROCK · NYUv2 val | mAP@0.5: 33.2 -> 37.6 | Improved |
| add depth supervision (ablation) | ROCK · NYUv2 val | mAP@0.5: 35.1 -> 37.6 | Improved |
| enable auxiliary losses (vs weights set to 0) (ablation) | ROCK · NYUv2 val | mAP@0.5: 30.6 -> 37.6 | Improved |
| element-wise product instead of addition for fusion (ablation) | ROCK depth-only · NYUv2 val | mAP@0.5: 30.9 -> 32.3 | Improved |
| pre-train on synthetic MLT dataset then fine-tune | ROCK (DNS) · NYUv2 test | mAP@0.5: 39.8 -> 46.8 | Improved |
| apply ROCK with disparity auxiliary task | SSD ResNet-50 · CityScapes val | mAP@0.5: 42.8 -> 44.0 | Improved |
| train on 70% of data with ROCK vs baseline on 100% | ROCK · NYUv2 val | mAP@0.5: 31.2 vs 32.8 | Mixed |
Closest papers by meaning
- Multi-Task Learning for Dense Prediction Tasks: A Survey (TPAMI 2021 · similarity 0.86)
- AutoMTL: A Programming Framework for Automating Efficient Multi-Task Learning (NeurIPS 2022 · similarity 0.86)
- Multi-Task Self-Supervised Object Detection via Recycling of Bounding Box Annotations (CVPR 2019 · similarity 0.85)
- CORE-MTL: Rethinking Gradient Balancing via Causal Orthogonal Representations (ICML 2026 · similarity 0.85)
- Roboflow100-VL: A Multi-Domain Object Detection Benchmark for Vision-Language Models (NeurIPS 2025 · similarity 0.83)
- Transforming Vision Transformer: Towards Efficient Multi-Task Asynchronous Learner (NeurIPS 2024 · similarity 0.83)
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.
