NeurIPS'19 Guided Similarity Separation for Image Retrieval

Authors: Chundi Liu, Guangwei Yu, Cheng Chang, Himanshu Rai, Junwei Ma, Satya Krishna Gorti, Maksims Volkovs
[paper][poster][slides][video]

Prerequisites and Environment

• tensorflow-gpu 1.13.1
• numpy 1.16.0

All experiments were conducted on a 20-core Intel(R) Xeon(R) CPU E5-2630 v4 @2.20GHz and NVIDIA V100 GPU with 32GB GPU memory.

Note: INSTRE dataset contains almost 30k images so the training phase requires ~19GB GPU memory. If your GPU doesn't have sufficient memory, please remove the --gpu-id argument and the model will train on the CPU. Training on CPU is approximately 10x slower but converges to the same result.

Dataset

We provide all three datasets used in the paper, together with generated GeM descriptors and RANSAC verification results for each dataset. To run the model, download the data from here and extract it to a directory, referred to as $DATA_DIR. This directory should have the following structure:

$DATA_DIR
  ├─ datasets
  │   ├─ instre
  |   |   ├─ gnd_instre.mat
  |   |   ├─ instre_gem.mat
  |   |   └─ instre_siamac.mat
  │   ├─ roxford5k
  |   |   ├─ gnd_roxford5k.mat
  |   |   └─ gnd_roxford5k.pkl
  |   └─ rparis6k
  |   |   ├─ gnd_rparis6k.mat
  |   |   └─ gnd_rparis6k.pkl
  ├─ features				
  │   ├─ instre_gem_index_ms_lw.npy
  │   ├─ instre_gem_query_ms_lw.npy
  │   ├─ roxford5k_resnet_rsfm120k_gem.mat
  │   └─ rparis6k_resnet_rsfm120k_gem.mat
  └─ graphs
      ├─ instre_index_ransac_graph.npy
      ├─ instre_query_ransac_graph.npy
      ├─ roxford5k_index_ransac_graph.npy
      ├─ roxford5k_query_ransac_graph.npy
      ├─ rparis6k_index_ransac_graph.npy
      └─ rparis6k_query_ransac_graph.npy

Provide $DATA_DIR as the argument to --data-path when running train.py.

Descriptors

For all the experiments, we use image descriptors generated by the pre-trained GeM model. The code and the pre-trained weights can be found in the author's official github repository.

Examples

---

rOxford medium 77.8

python train.py --data-path $DATA_DIR --dataset roxford5k --num-layers 2 --k 5 --kq 5 --epoch 200 --lr 0.0001 --gpu-id 0 --graph-mode descriptor --beta-percentile 98

rOxford medium ransac 80.6

python train.py --data-path $DATA_DIR --dataset roxford5k --num-layers 2 --k 5 --kq 10 --epoch 200 --lr 0.0001 --gpu-id 0 --graph-mode ransac --beta-percentile 98

rOxford hard 57.5

python train.py --data-path $DATA_DIR --dataset roxford5k --num-layers 2 --k 5 --kq 5 --epoch 200 --lr 0.0001 --gpu-id 0 --graph-mode descriptor --report-hard --beta-percentile 98

rOxford hard ransac 64.7

python train.py --data-path $DATA_DIR --dataset roxford5k --num-layers 2 --k 5 --kq 10 --epoch 200 --lr 0.0001 --gpu-id 0 --graph-mode ransac --report-hard --beta-percentile 98

rParis medium 92.4

python train.py --data-path $DATA_DIR --dataset rparis6k --num-layers 2 --k 5 --kq 15 --epoch 200 --lr 0.0001 --gpu-id 0 --graph-mode descriptor --beta-percentile 98

rParis medium ransac 93.4

python train.py --data-path $DATA_DIR --dataset rparis6k --num-layers 2 --k 5 --kq 25 --epoch 200 --lr 0.0001 --gpu-id 0 --graph-mode ransac --beta-percentile 98

rParis hard 83.5

python train.py --data-path $DATA_DIR --dataset rparis6k --num-layers 2 --k 5 --kq 15 --epoch 200 --lr 0.0001 --gpu-id 0 --graph-mode descriptor --report-hard --beta-percentile 98

rParis hard ransac 85.3

python train.py --data-path $DATA_DIR --dataset rparis6k --num-layers 2 --k 5 --kq 25 --epoch 200 --lr 0.0001 --gpu-id 0 --graph-mode ransac --report-hard --beta-percentile 98

INSTRE 89.2

python train.py --data-path $DATA_DIR --dataset instre --num-layers 2 --k 10 --kq 10 --epoch 500 --lr 0.0001 --graph-mode descriptor --beta-percentile 98

INSTRE ransac 92.4

python train.py --data-path $DATA_DIR --dataset instre --num-layers 2 --k 50 --kq 20 --epoch 500 --lr 0.0001 --graph-mode ransac --beta-percentile 98

Citation

If you find this code useful in your research, please cite the following paper:

@inproceedings{liu2019guided,
  title={Guided Similarity Separation for Image Retrieval},
  author={Chundi Liu, Guangwei Yu, Cheng Chang, Himanshu Rai, Junwei Ma, Satya Krishna Gorti, Maksims Volkovs},
  booktitle={NeurIPS},
  year={2019}
}