Research map / Architectures and efficiency / CNN design
Deep Supervision with Intermediate Concepts
Chi Li, M. Zeeshan Zia, Quoc-Huy Tran, Xiang Yu, Gregory D. Hager, Manmohan Krishna Chandraker
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 |
|---|---|---|---|
| supervise hidden layers with ordered intermediate concepts instead of plain single-task output (ablation) | DISCO · CIFAR100 | classification error (%): 23.31 -> 22.46 | Improved |
| order concepts by structure instead of random concept assignment (ablation) | DISCO · CIFAR100 | classification error (%): 27.53 -> 22.46 | Improved |
| supervise 3D shape concepts in intermediate layers rather than all in the output layer (plain-all) (ablation) | DISCO · KITTI-3D 2D (All) | PCK@0.1 (%): 80.6 -> 85.0 | Improved |
| supervise 3D shape concepts in intermediate layers rather than all in the output layer (plain-all) (ablation) | DISCO · KITTI-3D 3D (Full) | 3D PCK@0.1 (%): 92.9 -> 95.3 | Improved |
| add more concepts in the correct order: DISCO-3D-2D to DISCO-vis-3D-2D to DISCO (ablation) | DISCO · KITTI-3D 2D (All) | PCK@0.1 (%): 80.7 -> 83.4 -> 85.0 | Improved |
| use the proposed supervision order instead of reversed order (ablation) | DISCO · KITTI-3D 2D (All) | PCK@0.1 (%): 25.6 -> 85.0 | Improved |
| use the proposed order instead of 3D-vis alternative order (ablation) | DISCO · KITTI-3D 2D (All) | PCK@0.1 (%): 84.5 -> 85.0 | Improved |
| remove local spatial pooling and use global average pooling instead of VGG-style (ablation) | DISCO · KITTI-3D 2D (All) | PCK@0.1 (%): 69.0 -> 85.0 | Improved |
| add simulated truncation and multi-car occlusion to training data (ablation) | DISCO · KITTI-3D 2D (Truncation) | PCK@0.1 (%): 53.6 -> 78.5 | Improved |
| add simulated truncation and multi-car occlusion to training data (ablation) | DISCO · KITTI-3D 2D (Full) | PCK@0.1 (%): 91.8 -> 93.1 | Improved |
| use RCNN detection boxes instead of ground-truth boxes | DISCO · KITTI-3D 2D (All) | PCK@0.1 (%): 85.0 -> 88.3 | Improved |
| compare synthetic-only DISCO to real-trained VpsKps on small images | DISCO · PASCAL VOC car (Small Image) | PCK@0.1 (%): 67.4 -> 74.3 | Improved |
Closest papers by meaning
- Naturally Supervised 3D Visual Grounding with Language-Regularized Concept Learners (CVPR 2024 · similarity 0.88)
- Learning Good Features to Transfer Across Tasks and Domains (TPAMI 2023 · similarity 0.88)
- Deep Supervision With Shape Concepts for Occlusion-Aware 3D Object Parsing (CVPR 2017 · similarity 0.88)
- Attend, Infer, Repeat: Fast Scene Understanding with Generative Models (NeurIPS 2016 · similarity 0.87)
- InstructCV: Instruction-Tuned Text-to-Image Diffusion Models as Vision Generalists (ICLR 2024 · similarity 0.86)
- Towards flexible perception with visual memory (ICML 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.
