Research map / Detection and segmentation / CNN features
Compositional Convolutional Neural Networks: A Robust and Interpretable Model for Object Recognition Under Occlusion
Adam Kortylewski, Qing Liu, Angtian Wang, Yihong Sun, Alan Yuille
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 |
|---|---|---|---|
| replace fully connected head with compositional model (CompNet-RXT-RB4) | ResNext · Occluded-P3D+-Vehicles classification (mean over L0-L3) | accuracy: 83.4 -> 94.1 | Improved |
| add generative regularization of vMF kernels (gamma1=3) (ablation) | CompNet-ResNext-RB4 without regularization · Occluded-P3D+-Vehicles classification (mean over L0-L3) | accuracy: 84.4 -> 93.0 | Improved |
| add generative regularization of mixture components (gamma2=3) (ablation) | CompNet-ResNext-RB4 without regularization · Occluded-P3D+-Vehicles classification (mean over L0-L3) | accuracy: 84.4 -> 90.1 | Improved |
| add mixture-component regularization to vMF regularization (ablation) | CompNet-ResNext-RB4 · Occluded-P3D+-Vehicles classification (mean over L0-L3) | accuracy: 93.0 -> 94.1 | Improved |
| add vMF-kernel regularization to mixture regularization (ablation) | CompNet-ResNext-RB4 · Occluded-P3D+-Vehicles classification (mean over L0-L3) | accuracy: 90.1 -> 94.1 | Improved |
| end-to-end training with regularization after initialization (ablation) | CompNet-ResNext-RB4 · Occluded-P3D+-Vehicles classification (mean over L0-L3) | accuracy: 87.0 -> 94.1 | Improved |
| learn from RB4 instead of RB3 of ResNext (ablation) | CompositionalNet · Occluded-P3D+-Vehicles classification (mean over L0-L3) | accuracy: 88.8 -> 94.1 | Improved |
| use pool4 instead of pool5 of VGG (ablation) | CompNet-VGG · Occluded-P3D+-Vehicles classification (mean over L0-L3) | accuracy: 93.0 -> 92.2 | No clear effect |
| add CutPaste occlusion augmentation to training | ResNext · Occluded-COCO-Vehicles classification (avg L0-L3) | accuracy: 88.5 -> 94.1 | Improved |
| add CutPaste occlusion augmentation to training | CompNet-RXT-RB4 · Occluded-COCO-Vehicles classification (avg L0-L3) | accuracy: 94.8 -> 95.0 | No clear effect |
| reduce context weight from omega=0.5 to 0.2 (ablation) | CompNet-VGG-p4 · Occluded-P3D+-Vehicles detection (mean AP@IoU0.5) | AP@IoU0.5: 76.6 -> 81.3 | Improved |
| reduce context weight from omega=0.5 to 0.2 (ablation) | CompNet-VGG-p4 · Occluded-COCO-Vehicles detection (avg AP@IoU0.5) | AP@IoU0.5: 61.1 -> 72.4 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Faster R-CNN — Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks (TPAMI 2016) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Compositional Convolutional Neural Networks: A Deep Architecture With Innate Robustness to Partial Occlusion (CVPR 2020 · similarity 0.97)
- Robust Object Detection Under Occlusion With Context-Aware CompositionalNets (CVPR 2020 · similarity 0.95)
- Teaching Compositionality to CNNs (CVPR 2017 · similarity 0.90)
- Increasing CNN Robustness to Occlusions by Reducing Filter Support (ICCV 2017 · similarity 0.90)
- A Bayesian Approach to OOD Robustness in Image Classification (CVPR 2024 · similarity 0.88)
- Multi-Part Object Representations via Graph Structures and Co-Part Discovery (NeurIPS 2026 · similarity 0.87)
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.
