DELG-pytorch

Pytorch Implementation of Unifying Deep Local and Global Features for Image Search (delg-eccv20)

• DELG pipline:

Installation

Install Python dependencies:

pip install -r requirements.txt

Set PYTHONPATH:

export PYTHONPATH=`pwd`:$PYTHONPATH

Training

Training a delg model:

python train_delg.py \
    --cfg configs/metric/resnet_delg_8gpu.yaml \
    OUT_DIR ./output \
    PORT 12001 \
    TRAIN.WEIGHTS path/to/pretrainedmodel

Resume training:

python train_delg.py \
    --cfg configs/metric/resnet_delg_8gpu.yaml \
    OUT_DIR ./output \
    PORT 12001 \
    TRAIN.AUTO_RESUME True

Weights

-r50-delg (wu46)

-r101-delg (5pdj)

pretrained weeights are available in pymetric

Feature extraction

!!! Queries should be cropped as DOLG.

Extracting global and local feature for multi-scales

python tools/extractor.py --cfg configs/resnet_delg_8gpu.yaml

Refer extractor.sh for using multicards

See visualize.ipynb for verification of local features

Evaluation on ROxf and RPar

Local Match

• Spatial Verification

  Install pydegensac and see tools/rerank/spatial_verification.py

• Examples

• ASMK

  (https://github.com/jenicek/asmk)

Results

See (https://github.com/filipradenovic/revisitop) for details

cd tools/revisitop
python example_evaluate_with_local.py main

• on roxford5k

Backbone	Train Size	Method	mAP E	mAP M	mAP H
ResNet50	224	Global Ranking	77.73	66.06	38.37
ResNet50	224	Global	81.03	68.31	39.98
ResNet50	224	Global + Spatial Verification	84.81	71.97	46.63
ResNet50	512	Global	90.55	78.51	56.90
ResNet50	512	Global + Spatial Verification	90.86	80.08	58.42

• on rparis6k(updating)

1. SOTA of R50-DELG is 78.3 mAP@M in the paper, we outperform it
2. All training set version is GLDv2-clean (81313, 1580470)
3. Traing size, global and local feature scales adopted are same with the paper