InsCLR: Improving Instance Retrieval with Self-Supervision

This is an official PyTorch implementation of the InsCLR paper.

Download Dataset

Dataset	Image	Annotation
The Oxford Buildings Dataset	download	download
The Paris Dataset	download	download
Google Landmarks Dataset v2	download	download
R-1M distractor	download
INSTRE	download	download

We also provide scripts for downloading these datasets (see download).

Training

To meet the performance reported in the paper, you need several training stages, and each training stage may have a different config, but they share a common pipeline.

Generally, a training stage includes the following steps.

Extract Features

Using backbone pretrained on ImageNet or trained on previous stage to extract features of Google Landmarks Dataset v2 or INSTRE.

export CUDA_VISIBLE_DEVICES=0,1,2,3
export PYTHONPATH=$PWD
python3 tools/compute_candidate_pool.py --cfg configs/instre/base.yaml --task feature --dataset instre --name s0_r50 --pretrain imagenet

Compute Candidate Pool for Each Image

As mentioned in Section 3.1: Setup for training samples.

export CUDA_VISIBLE_DEVICES=0,1,2,3
export PYTHONPATH=$PWD
python3 tools/compute_candidate_pool.py --task neighbor --dataset instre --name s0_r50

Configuration

There are two different configs we use in the whole training, we call them base and impr config, respectively.

For Google Landmarks Dataset v2, We use base config for the first three training stages, and impr config in the fourth training stage, so the whole training contains four stages. And for INSTRE, only one base config training stage is considered.

Start Training

export CUDA_VISIBLE_DEVICES=0,1,2,3
export PYTHONPATH=$PWD
tools/train_net.py --cfg /path/to/config

justforforfor/insclr