# AffNet

> CNN-based affine shape estimator implemented in PyTorch for learning discriminative affine regions via discriminability, as presented at ECCV 2018.

AffNet is a CNN-based affine shape estimator implemented in PyTorch, released alongside the ECCV 2018 paper "Repeatability Is Not Enough: Learning Discriminative Affine Regions via Discriminability" by Dmytro Mishkin, Filip Radenovic, and Jiri Matas. The project provides code, pre-trained weights, and usage examples for estimating affine shapes of local image features, and is available under the MIT License on GitHub.

## What It Is

AffNet is a deep learning model that learns to estimate the affine shape of local image regions, replacing classical iterative methods like Baumberg iterations with a neural network approach. It is designed to work alongside local feature detectors (such as Hessian-Affine) and descriptors (such as HardNet++) to improve image matching and retrieval performance. The core contribution is training the network for discriminability rather than repeatability alone, which the authors argue leads to better correspondences in practice.

## Performance on Image Retrieval

The repository benchmarks AffNet against classical baselines on the Oxford5k dataset using mean Average Precision (mAP). According to the README, combining HesAffNet with HardNet++ achieves mAP scores of 68.3 (BoW), 77.8 (BoW + SV), 89.0 (BoW + SV + QE), and 89.5 (HQE + MA), outperforming HesAff + RootSIFT and HesAff + HardNet++ across all evaluation settings. The README also states that AffNet generates up to twice more correspondences compared to Baumberg iterations.

## Architecture and Usage

AffNet is a convolutional neural network trained in PyTorch. The repository includes:
- Pre-trained model weights (`AffNet.pth`) in the `pretrained` folder
- Example scripts for estimating affine shape on patch-column files in HPatches format
- A PyTorch implementation of Hessian-Affine using AffNet for region detection
- Jupyter notebooks for reproducing paper figures and a WBS (Wide Baseline Stereo) demo
- Output in Oxford affine format compatible with standard benchmarks

The input/output interface supports grayscale patch images and produces affine frame parameters (a11, a21, a22) or Oxford affine ellipse format.

## Update: PyTorch 1.4 Branch

The master branch corresponds to the original ECCV paper results, targeting Python 2.7 and PyTorch 0.4.0. A separate branch (`pytorch1-4_python3`) is available for users running Python 3.7 and PyTorch 1.4.0, making the codebase accessible on more modern environments. The repository was last pushed to in April 2025, indicating ongoing maintenance activity.

## Audience and Deployment

AffNet targets computer vision researchers and practitioners working on image matching, image retrieval, and local feature pipelines. It is deployed as source code cloned from GitHub, with a shell script (`run_me.sh`) to download datasets and begin training. The project is cited in academic work and is intended for use as a component within larger feature extraction and matching pipelines.

## Features
- CNN-based affine shape estimation
- Pre-trained AffNet.pth model weights
- PyTorch implementation of Hessian-Affine detector
- HPatches format patch input support
- Oxford affine ellipse format output
- Jupyter notebook examples for paper figure reproduction
- WBS (Wide Baseline Stereo) demo notebook
- Comparison with Baumberg iterations
- Dataset download and training scripts
- Python 3.7 / PyTorch 1.4 compatible branch

## Integrations
PyTorch, HardNet++, HesAff, HPatches benchmark, Oxford affine benchmark

## Platforms
API, DEVELOPER_SDK, CLI

## Pricing
Open Source

## Version
pytorch1-4_python3 branch

## Links
- Website: https://github.com/ducha-aiki/affnet
- Documentation: https://github.com/ducha-aiki/affnet
- Repository: https://github.com/ducha-aiki/affnet
- EveryDev.ai: https://www.everydev.ai/tools/affnet