Research map / Detection and segmentation / Weakly supervised detection and segmentation
CAM Back Again: Large Kernel CNNs from a Weakly Supervised Object Localization Perspective
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Revisits large kernel CNNs (ConvNeXt, RepLKNet, SLaK) on weakly supervised object localization with CAM, finding feature map quality rather than effective receptive field size drives WSOL gains, and reaching 90.99% MaxBoxAcc on CUB-200-2011.
Problem
Whether the high downstream performance of large kernel CNNs stems from large effective receptive fields (ERF), tested via weakly supervised object localization (WSOL) on CUB-200-2011, where CAM tends to activate only local object parts.
Approach
Fine-tune modern large kernel CNNs and apply plain CAM to show that architecture-driven feature map quality (globally activated, cleaner maps, large GAP/activation areas) largely explains WSOL gains, not ERF size; a PCA-based PC1 localization map further improves scores.
Key results
- CAM with fine-tuned RepLKNet-31B1K384 reaches 90.99% MaxBoxAcc on CUB-200-2011, comparable to state-of-the-art CNN-based WSOL methods.
- A PCA-based PC1 localization map with RepLKNet achieves 93.6% MaxBoxAcc, reported as 0.43% above the prior CNN-based state of the art.
- ERF size shows only an unstable relationship with WSOL score; SLaK's ERF is often smaller than RepLKNet's despite larger kernels, and data augmentation changes WSOL scores by -5.8% to +4.0%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use binarized PC1 features instead of CAM maps | RepLKNet · CUB-200-2011 | MaxBoxAcc: 93.6% ; 0.43% higher than the state-of-the-art performance of CNN-based WSOL | Helped |
| Add data augmentation during fine-tuning (ablation) | RepLKNet 31B1K 224px · CUB-200-2011 | MaxBoxAcc: 77.68 -> 74.39; -3.30 | Hurt |
| Add data augmentation during fine-tuning (ablation) | RepLKNet 31B1K 384px · CUB-200-2011 | MaxBoxAcc: 85.30 -> 87.47; +2.17 | Helped |
| Add data augmentation during fine-tuning (ablation) | RepLKNet 31B22K 224px · CUB-200-2011 | MaxBoxAcc: 64.07 -> 60.86; -3.21 | Hurt |
| Add data augmentation during fine-tuning (ablation) | RepLKNet 31B22K 384px · CUB-200-2011 | MaxBoxAcc: 67.73 -> 68.31; +0.59 | Helped |
| Add data augmentation during fine-tuning (ablation) | RepLKNet 31L22K 384px · CUB-200-2011 | MaxBoxAcc: 55.59 -> 59.58; +3.99 | Helped |
| Add data augmentation during fine-tuning (ablation) | RepLKNet XL MegData 320px · CUB-200-2011 | MaxBoxAcc: 73.35 -> 67.59; -5.76 | Hurt |
| Use RepLKNet as the backbone for CAM | CAM · CUB-200-2011 | MaxBoxAccV1: 90.99; 90.3 | Helped |
Surprising findings
- RepLKNet pretrained without data augmentation scored higher than 7 of 11 models trained with augmentation, suggesting architecture matters more than augmentation.
- SLaK with larger kernels produced more locally activated CAMs than RepLKNet and ConvNeXt, contrary to the expectation that larger kernels help.
Limitations
- Results are reported only on CUB-200-2011 for the main claims, with a single WSOL task.
- The best score relies on checkpoint selection and threshold choices using the ImageNet fullsup split, with some baselines marked as not using separate thresholds.
- Mechanistic explanation for why ConvNeXt and RepLKNet produce global feature maps is not established.
Experimental setup
Datasets: ImageNet1K CUB-200-2011
Benchmarks: CUB-200-2011 ImageNet1K
Metrics: MaxBoxAcc MaxBoxAccV2 MacBoxAcc ERF size (AUC) IoU threshold
Baselines: CAM HaS ACoL DDT F-CAM CREAM BridgeGAP
Builds on: CAM RepLKNet SLaK ConvNeXt CutMix
Closest papers by meaning
- More ConvNets in the 2020s: Scaling up Kernels Beyond 51x51 using Sparsity (ICLR 2023 · similarity 0.89)
- Scaling Up Your Kernels to 31x31: Revisiting Large Kernel Design in CNNs (CVPR 2022 · similarity 0.89)
- Recognize Any Regions (NeurIPS 2024 · similarity 0.86)
- UniRepLKNet: A Universal Perception Large-Kernel ConvNet for Audio Video Point Cloud Time-Series and Image Recognition (CVPR 2024 · similarity 0.86)
- Role of Locality and Weight Sharing in Image-Based Tasks: A Sample Complexity Separation between CNNs, LCNs, and FCNs (ICLR 2024 · similarity 0.86)
- PeLK: Parameter-efficient Large Kernel ConvNets with Peripheral Convolution (CVPR 2024 · 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.