Pytorch Implementation of Deep Local Feature (DeLF)

PyTorch Implementation of "Large-Scale Image Retrieval with Attentive Deep Local Features"
reference: https://arxiv.org/pdf/1612.06321.pdf

Prerequisites

• PyTorch
• python3
• CUDA

Training DeLF

There are 2 steps for DeLF training: (1) finetune stage, and (2) keypoint stage.
Finetune stage loads resnet50 model pretrained on ImageNet, and finetune.
Keypoint stage freezes the "base" network, and only update "attention" network for keypoint selection. After the train process is done, model will be saved at repo/<expr>/keypoint/ckpt

(1) training finetune stage:

$ cd train/
$ python main.py \
    --stage 'finetune' \
    --optim 'sgd' \
    --gpu_id 6 \
    --expr 'landmark' \
    --ncls 586 \
    --finetune_train_path <path to train data> \
    --finetune_val_path <path to val data> \

(2) training keypoint stage:

• load_from: absolute path to pytorch model you wish to load. (<model_name>.pth.tar)
• expr: name of experiment you wish to save as.

$ cd train/
$ python main.py \
    --stage 'keypoint' \
    --gpu_id 6 \
    --ncls 586 \
    --optim 'sgd' \
    --use_random_gamma_scaling true \
    --expr 'landmark' \
    --load_from <path to model> \
    --keypoint_train_path <path to train data> \
    --keypoint_val_path <path to val data> \

Feature Extraction of DeLF

There are also two steps to extract DeLF: (1) train PCA, (2) extract dimension reduced DeLF.
IMPORTANT: YOU MUST CHANGE OR COPY THE NAME OF MODEL from repo/<expr>/keypoint/ckpt/bestshot.pth.tar to repo/<expr>/keypoint/ckpt/fix.pth.tar.
I intentionally added this to prevent the model from being updated after the PCA matrix is already calculated.

(1) train PCA

$ cd extract/
$ python extractor.py
    --gpu_id 4 \
    --load_expr 'delf' \
    --mode 'pca' \
    --stage 'inference' \
    --batch_size 1 \
    --input_path <path to train data>, but it is hardcoded.
    --output_path <output path to save pca matrix>, but it is hardcoded.

(2) extract dimension reduced DeLF

$ cd extract/
$ python extractor.py
    --gpu_id 4 \
    --load_expr 'delf' \
    --mode 'delf' \
    --stage 'inference' \
    --batch_size 1 \
    --attn_thres 0.31 \
    --iou_thres 0.92 \
    --top_k 1000 \
    --use_pca True \
    --pca_dims 40 \
    --pca_parameters_path <path to pca matrix file.>, but it is hardcoded.
    --input_path <path to train data>, but it is hardcoded.
    --output_path <output path to save pca matrix>, but it is hardcoded.

Visualization

You can visualize DeLF matching batween two arbitrary query images. Let's assume there exist two images, test/img1.jpg, test/img2.jpg. Run visualize.ipynb using Jupyter Notebook, and run each cells. You may get the result like below.

1) RANSAC Matching (Correspondance Matching + Geometric Verification)

2) Attention Map

Ranking Result on Oxf5k:

• glr1k, glr2k: Trained DeLF model with a subset of google-landmark-dataset on kaggle, which contains top-K instances sorted by the # of images included.
  ** ldmk: Trained DeLF model with landmark dataset. (exactly same with the paper)

glr1k ranking result

glr2k ranking result

ldmk ranking result

Benchmark Result on Oxf5k (comparing to original paper)

Note: DELF_TF is the author's model, and the feature was extracted using this nice repo. (https://github.com/insikk/delf_enhanced)
PYTORCH_LDMK: Trained with landmark dataset.
PYTORCH_GLR1K: Trained with a subset of google-landmark-dataset with 1k instance classes.
PYTORCH_GLR1K: Trained with a subset of google-landmark-dataset with 2k instance classes.
PYTORCH_BNK_V3_BAL_HANA: Private currency dataset I personally own just for check.

Classes	DELF_TF	PYTORCH_LDMK	PYTORCH_GLR1K	PYTORCH_GLR2K	PYTORCH_BNK_V3_BAL_HANA
mAP	0.851307	0.849373	0.87828	0.866517	0.489614
all_souls_1	0.751052	0.767453	0.916059	0.886243	0.0584418
all_souls_2	0.517995	0.645628	0.708546	0.767904	0.287783
all_souls_3	0.626499	0.760189	0.881578	0.903977	0.347261
all_souls_4	0.968566	0.930445	0.967221	0.980288	0.515091
all_souls_5	0.735256	0.827341	0.899803	0.911414	0.117378
ashmolean_1	0.83206	0.768585	0.829522	0.860364	0.157126
ashmolean_2	0.844329	0.803305	0.814522	0.88631	0.194069
ashmolean_3	0.8407	0.863916	0.86428	0.841624	0.20158
ashmolean_4	0.857416	0.730968	0.816007	0.829129	0.353456
ashmolean_5	0.77901	0.84768	0.808717	0.875755	0.106619
balliol_1	0.917435	0.818512	0.914453	0.857404	0.362258
balliol_2	0.462124	0.5546	0.68825	0.632167	0.0984046
balliol_3	0.710849	0.72742	0.80883	0.729275	0.209934
balliol_4	0.658099	0.681549	0.749764	0.667446	0.342497
balliol_5	0.739436	0.689549	0.80835	0.716029	0.319832
bodleian_1	0.7943	0.797353	0.833887	0.851872	0.350422
bodleian_2	0.828246	0.549165	0.520681	0.413119	0.643002
bodleian_3	0.84655	0.844758	0.954003	0.841856	0.799652
bodleian_4	0.726362	0.732197	0.916468	0.84604	0.476852
bodleian_5	0.815629	0.864863	0.915992	0.847784	0.773505
christ_church_1	0.953197	0.97743	0.96955	0.987822	0.866622
christ_church_2	0.960692	0.950959	0.975525	0.979186	0.783949
christ_church_3	0.932694	0.951987	0.940492	0.942081	0.263114
christ_church_4	0.965374	0.979779	0.970264	0.981529	0.784185
christ_church_5	0.971503	0.971411	0.976488	0.983004	0.312071
cornmarket_1	0.690551	0.722799	0.692261	0.681911	0.492891
cornmarket_2	0.727338	0.382168	0.32282	0.184599	0.169908
cornmarket_3	0.707911	0.650324	0.696718	0.672553	0.379656
cornmarket_4	0.65958	0.789562	0.656362	0.669228	0.273514
cornmarket_5	0.68901	0.814039	0.606983	0.558519	0.19587
hertford_1	0.92893	0.915811	0.957557	0.951947	0.562145
hertford_2	0.960313	0.942536	0.937546	0.951293	0.524951
hertford_3	0.936073	0.959108	0.97494	0.941641	0.570177
hertford_4	0.898146	0.914434	0.924889	0.927225	0.679879
hertford_5	0.975377	0.929499	0.946097	0.94726	0.235865
keble_1	1	1	1	1	0.954762
keble_2	1	0.944161	1	1	0.921088
keble_3	1	0.932568	1	1	0.931319
keble_4	1	1	1	1	0.331796
keble_5	1	0.87432	1	1	0.944161
magdalen_1	0.710288	0.766209	0.819577	0.861361	0.109972
magdalen_2	0.830566	0.928487	0.914451	0.926896	0.164253
magdalen_3	0.759041	0.832379	0.872577	0.896532	0.168931
magdalen_4	0.853145	0.877747	0.880979	0.844535	0.0728258
magdalen_5	0.761443	0.77776	0.841862	0.791102	0.175314
pitt_rivers_1	1	1	1	1	0.647935
pitt_rivers_2	1	1	1	1	1
pitt_rivers_3	1	1	1	1	0.746479
pitt_rivers_4	1	1	1	1	0.599398
pitt_rivers_5	1	1	1	1	1
radcliffe_camera_1	0.93144	0.916562	0.943584	0.95298	0.860801
radcliffe_camera_2	0.961224	0.980161	0.980304	0.982237	0.936467
radcliffe_camera_3	0.925759	0.908404	0.949748	0.959252	0.871228
radcliffe_camera_4	0.979608	0.98273	0.983941	0.988227	0.787773
radcliffe_camera_5	0.90082	0.936742	0.952967	0.949522	0.894346

Author

Minchul Shin(@nashory)
contact: min.nashory@navercorp.com