This repository provides a PyTorch implementation of Bilinear Convolutional Neural Networks (B-CNN) for fine-grained image recognition. It addresses the need for richer image representations by capturing pairwise descriptor interactions, outperforming part-based models without requiring explicit part annotations. The target audience includes researchers and practitioners in computer vision focused on fine-grained classification tasks.

How It Works

B-CNN employs bilinear pooling to compute the sum of outer products of deep image descriptors. This approach captures all pairwise descriptor interactions in a translation-invariant manner, yielding richer representations than linear models. The implementation allows for fine-tuning either just the fully connected layer or all layers of the network.

Quick Start & Requirements

• Install: Requires Python 3.6 with NumPy and PyTorch.
• Usage:
  • Fine-tune FC layer only: ./src/bilinear_cnn_fc.py --base_lr 1.0 --batch_size 64 --epochs 55 --weight_decay 1e-8
  • Fine-tune all layers: ./src/bilinear_cnn_all.py --base_lr 1e-2 --batch_size 64 --epochs 25 --weight_decay 1e-5 --model "model.pth"
• Hardware: Supports multi-GPU training (e.g., CUDA_VISIBLE_DEVICES=0,1,2,3).
• Links: Original Paper

Highlighted Details

• Achieves 76.77% test set accuracy when fine-tuning only the FC layer.
• Achieves 84.17% test set accuracy when fine-tuning all layers.
• Captures pairwise descriptor interactions for improved representation richness.

Maintenance & Community

• The project is relatively old, written for PyTorch 0.3.0.
• A faster alternative for newer PyTorch versions is available at HaoMood/blinear-cnn-faster.
• Author: Hao Zhang (zhangh0214@gmail.com).

Licensing & Compatibility

• License: CC BY-SA 3.0.
• This license is a Creative Commons Attribution-ShareAlike license, which requires derivative works to be shared under the same or a compatible license. Commercial use may be restricted depending on how the code is integrated and distributed.

Limitations & Caveats

The repository is written for an outdated PyTorch version (0.3.0), necessitating potential migration efforts for compatibility with modern PyTorch releases.