🌉 Beyond Softmax: Dual-Branch Sigmoid Architecture for Accurate Class Activation Maps (BMVC2025)

This repository contains official code implementation of "Beyond Softmax: Dual-Branch Sigmoid Architecture for Accurate Class Activation Maps". We provide PyTorch training, inference code and hyperparameter settings reported in the paper.

Abstract

In this work, we address two fundamental distortions in Class Activation Mapping (CAM) and its variants—additive logit shifts and sign collapse—both arising from reliance on a final softmax classifier. We propose a simple, architecture-agnostic dual-branch sigmoid head that decouples localization from classification. Our method fine-tunes a sigmoid branch in parallel with the frozen softmax head, enabling class evidence maps that preserve both magnitude and sign of feature contributions, while retaining original recognition accuracy. Experiments on fine-grained and WSOL benchmarks demonstrate improved explanation fidelity and consistent Top-1 Localization gains without sacrificing classification performance.

Getting started

This code is heavily borrowed from Evaluating Weakly Supervised Object Localization Methods Right and pytorch-grad-cam. The training settings for the backbones follow the shared setting spreadsheet provided from wsolevaluation.

1) Clone the repository

git clone beyond-softmax
cd beyond-softmax

2) Download datasets (ImageNet1K, CUB-200-2011, OpenImages30K, Stanford_Cars) then align with metadata/.

mkdir dataset
mv {DOWNLOADED_DATASET} ./dataset/

{DOWNLOADED_DATASET} is the path of folder that you downloaded dataset to.

3) Environment setup - CUDA 12.2 with RTX A6000 GPU

conda create -n dbs python=3.12 -y
conda activate dbs
pip install -r requirements.txt

4) Train

$CKPT_PATH should point to a checkpoint of the vanilla backbone model (e.g. Vanilla VGG-16) and available values for $METHOD are [cam, gradcam].

An example command line for the train:

python main.py \
  --experiment_name ${EXPERIMENT}$ \
  --dataset_name ILSVRC \
  --architecture vgg16 \
  --method ${METHOD} \
  --root ${ROOT} \
  --large_feature_map FALSE \
  --epoch 3 \
  --batch_size 32 \
  --lr_decay_frequency 3 \
  --workers 4 \
  --gpus 0 \
  --lr 0.003 \
  --weight_decay 5.00E-04 \
  --model_structure b2 \
  --unfreeze_layer fc2 \
  --ft_ckpt $CKPT_PATH 

5) Eval

Here, $METHOD can be [cam, gradcam, gradcam++, xgradcam, layercam].
An example command line for the test:

python main.py \
  --experiment_name ${EXPERIMENT}$ \
  --dataset_name ILSVRC \
  --architecture vgg16 \
  --method ${METHOD} \
  --root ${ROOT} \
  --large_feature_map FALSE \
  --batch_size 32 \
  --workers 4 \
  --gpus 0 \
  --model_structure b2 \
  --ft_ckpt $CKPT_PATH \
  --eval_only \
  --eval_checkpoint_type last 

Citations

If you find our work interesting, please consider citing

@article{oh:2025:beyondsoftmax,
  title = {Beyond Softmax: Dual-Branch Sigmoid Architecture for Accurate Class Activation Maps},
  author = {Oh, Yoojin and Noh, Junhyug},
  journal = {British Machine Vision Conference (BMVC)},
  year = {2025}
}