Research map / Detection and segmentation / Open-set and few-shot detection
Increasing CNN Robustness to Occlusions by Reducing Filter Support
Elad Osherov, Michael Lindenbaum
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 second-central-moment regularizer on conv filters | LeNet trained on unoccluded images · CIFAR-10 (LeNet, occluded val, delta=0.1) | error rate (%): 27.31 -> 23.47 | Improved |
| add group-sparsity regularizer on conv filters | LeNet trained on unoccluded images · CIFAR-10 (LeNet, occluded val, delta=0.1) | error rate (%): 27.31 -> 24.20 | Improved |
| add mutually re-weighted L1 regularizer on conv filters | LeNet trained on unoccluded images · CIFAR-100 (LeNet, occluded val, delta=0.1) | error rate (%): 60.17 -> 58.68 | Improved |
| add plain L1 regularization on conv+fc | LeNet trained on unoccluded images · CIFAR-10 (LeNet, occluded val, delta=0.1) | error rate (%): 27.31 -> 27.52 | No clear effect |
| add iterative reweighted L1 on conv+fc | LeNet trained on unoccluded images · CIFAR-10 (LeNet, occluded val, delta=0.1) | error rate (%): 27.31 -> 26.45 | Improved |
| also regularize first fully connected layer with second-central-moment term (ablation) | RCM conv · CIFAR-10 (LeNet, occluded val, delta=0.1) | error rate (%): 23.47 -> 23.00 | Improved |
| also regularize first fully connected layer with mutually re-weighted L1 (ablation) | RMRWL1 conv · CIFAR-100 (LeNet, occluded val, delta=0.1) | error rate (%): 58.68 -> 59.78 | Hurt |
| train on occluded examples instead of unoccluded | LeNet · CIFAR-10 (LeNet, occluded val, delta=0.1) | error rate (%): 27.31 -> 22.89 | Improved |
| RCM regularization vs training on occluded examples | LeNet · CIFAR-100 (LeNet, occluded val, delta=0.1) | error rate (%): 56.43 -> 56.22 | Improved |
| use strong lambda tuned for occlusion, tested on unoccluded images | RCM · CIFAR (LeNet, unoccluded val) | error rate (%): increases by 0.4% | Hurt |
| RCM with lambda tuned for unoccluded images instead of L2 | RCM · CIFAR (LeNet, unoccluded val) | error rate (%): goes down by 1.7% | Improved |
| add support regularization to VGG16 instead of occluded training | VGG16 · ILSVRC12 (occluded val, delta=0.1) | top-5 error (%): 16.9 -> 13.3 | Improved |
Closest papers by meaning
- Compositional Convolutional Neural Networks: A Deep Architecture With Innate Robustness to Partial Occlusion (CVPR 2020 · similarity 0.90)
- Compositional Convolutional Neural Networks: A Robust and Interpretable Model for Object Recognition Under Occlusion (IJCV 2020 · similarity 0.90)
- Robust Instance Segmentation Through Reasoning About Multi-Object Occlusion (CVPR 2021 · similarity 0.89)
- Robust Object Detection Under Occlusion With Context-Aware CompositionalNets (CVPR 2020 · similarity 0.89)
- A-Fast-RCNN: Hard Positive Generation via Adversary for Object Detection (CVPR 2017 · similarity 0.87)
- A Bayesian Approach to OOD Robustness in Image Classification (CVPR 2024 · 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.
