Reproduce “CAM Back Again: Large Kernel CNNs from a Weakly Supervised Object Localization Perspective” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Shunsuke Yasuki, Masato Taki

Revisiting classic Class Activation Map (CAM) with modern large kernel CNNs reveals that these architectures produce high-quality, globally activated feature maps. This enables CAM to achieve competitive performance in Weakly Supervised Object Localization (WSOL), with a new, simpler PC1-based method surpassing previous state-of-the-art results by 0.43% MaxBoxAcc on CUB-200-2011.

Key insights

  • The primary factor driving the performance of modern large kernel CNNs in WSOL is the improved quality of their feature maps, not solely their Effective Receptive Field (ERF) size.
  • Classic Class Activation Map (CAM)'s "partial localization" issue is largely mitigated by the globally activated feature maps produced by modern CNN backbones like ConvNeXt and RepLKNet.
  • Modern large kernel CNN architectures possess inherent biases, even before training, that promote the generation of more globally representative feature maps suitable for localization.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.