Research map / Detection and segmentation / Open-set and few-shot detection
A Random CNN Sees Objects: One Inductive Bias of CNN and Its Applications
Yun-Hao Cao, Jianxin Wu
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 Tobias-guided background replacement to MoCov2 pretraining | ResNet-50 on CUB-200 · CUB-200, Mixup fine-tuning, 200 pretraining epochs | accuracy (%): 53.0 -> 62.0 (+9.0) | Improved |
| Add Tobias-guided background replacement to SimCLR pretraining | ResNet-50 on CUB-200 · CUB-200, Mixup fine-tuning, 200 pretraining epochs | accuracy (%): 66.5 -> 71.7 (+5.2) | Improved |
| Add Tobias-guided background replacement to MoCov2 pretraining | ResNet-50 pretrained on ImageNet · Pascal VOC trainval07+12, R-50-FPN | AP75: 60.5 -> 61.1 | Improved |
| Add Tobias-guided background replacement to MoCov2 pretraining | ResNet-50 on IN-100k · VOC 07&12, 200 epochs | AP75: 51.6 -> 53.9 (+2.3) | Improved |
| Remove Tobias foreground/background guidance and merge patches randomly (ablation) | MoCov2-Tobias on IN-10k · VOC 07&12, 800 epochs | AP50: 73.2 -> 72.0 (↓1.2) | Hurt |
| Replace Tobias-guided merging with Mixup (ablation) | MoCov2-Tobias on IN-10k · VOC 07&12, 800 epochs | AP50: 73.2 -> 70.9 (↓2.3) | Hurt |
| Increase background replacement probability to p=0.5 (ablation) | MoCov2-Tobias on IN-10k · VOC 07&12, 800 epochs | AP75: 45.9 -> 49.4 | Improved |
| Set background replacement probability to p=0.7 (ablation) | MoCov2-Tobias on IN-10k · Tiny-IN-200, 800 epochs | linear evaluation accuracy (%): 23.6 -> 25.4 | Improved |
| Reduce ResNet-50 depth to bottleneck configuration [1,2,3,1] (ablation) | Random ResNet-50 localization · ImageNet | localization accuracy (%): 48.2±0.6 -> 43.9±1.5 | Hurt |
| Increase ResNet-50 depth to bottleneck configuration [6,8,12,6] (ablation) | Random ResNet-50 localization · CUB-200 | localization accuracy (%): 41.8±0.6 -> 45.0±0.9 | Improved |
| Replace ReLU with sigmoid (ablation) | Random ResNet-50 localization · CUB-200 | localization accuracy (%): 41.8±0.6 -> 22.6±3.3 | Hurt |
| Remove skip connections (ablation) | Random ResNet-50 localization · ImageNet | localization accuracy (%): 48.2±0.6 -> 50.8±1.0 | No clear effect |
Closest papers by meaning
- Tobias: A Random CNN Sees Objects (TPAMI 2023 · similarity 0.98)
- Spatially Invariant Unsupervised Object Detection with Convolutional Neural Networks (AAAI 2019 · similarity 0.86)
- Training Object Detectors From Scratch: An Empirical Study in the Era of Vision Transformer (CVPR 2022 · similarity 0.86)
- Label-Focused Inductive Bias over Latent Object Features in Visual Classification (ICLR 2024 · similarity 0.86)
- Training the Untrainable: Introducing Inductive Bias via Representational Alignment (NeurIPS 2025 · similarity 0.84)
- Grounding Everything: Emerging Localization Properties in Vision-Language Transformers (CVPR 2024 · similarity 0.84)
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.
