Hard_Negative

This repo is for paper in ECCV2020: Hard negative examples are hard, but useful

Arxiv link: https://arxiv.org/pdf/2007.12749.pdf

We provide the experiment result for each dataset(DATA) in folder exp_{DATA}.

DATA = CUB or CAR The _result folder contains tensorboard results with Keys: {DATA}_test_R_1, {DATA}_test_R_2, {DATA}_test_R_4, {DATA}_test_R_8, {DATA}_train_R_1, hn_ratio, loss.

DATA = SOP The _result folder contains tensorboard results with Keys: SOP_test_R_1, SOP_test_R_1, SOP_test_R_100, SOP_test_R_1000, SOP_train_R_1, hn_ratio, loss.

DATA = ICR(Inshop) The _result folder contains tensorboard results with Keys: ICR_test_R_1, ICR_test_R_1, ICR_test_R_100, ICR_test_R_1000, hn_ratio, loss.

Result Table

CUB(ResNet50, embedding size 64)

Method	R@1	R@2	R@4	R@8
hn	collapse	collapse	collapse	collapse
shn	56.72\pm0.65	68.64\pm0.33	78.60\pm0.24	86.64\pm0.26
sct	57.71\pm0.75	69.80\pm0.41	79.59\pm0.58	87.05\pm0.38

CAR(ResNet50, embedding size 64)

Method	R@1	R@2	R@4	R@8
hn	collapse	collapse	collapse	collapse
shn	67.92\pm0.49	77.80\pm0.39	85.31\pm0.18	90.69\pm0.14
sct	73.35\pm0.54	81.98\pm0.25	88.02\pm0.18	92.36\pm0.26

SOP(ResNet50, embedding size 512)

Method	R@1	R@10	R@100	R@1000
hn	not measured	not measured	not measured	not measured
shn	81.06\pm0.06	92.32\pm0.07	96.84\pm0.05	98.90\pm0.01
sct	81.90\pm0.07	92.61\pm0.06	96.77\pm0.04	98.75\pm0.02

Inshop(ResNet50, embedding size 512)

Method	R@1	R@10	R@20	R@40
hn	not measured	not measured	not measured	not measured
shn	90.55\pm0.15	97.37\pm0.07	98.09\pm0.10	98.45\pm0.08
sct	90.93\pm0.22	97.51\pm0.05	98.16\pm0.03	98.44\pm0.03

Hotel50K-instance(ResNet50, embedding size 256)

Method	R@1	R@10	R@100
hn	collapse	collapse	collapse
shn	18.78\pm0.08	32.90\pm0.26	52.19\pm0.19
sct	29.24\pm0.12	44.38\pm0.11	61.53\pm0.08

Hotel50K-chain(ResNet50, embedding size 256)

Method	R@1	R@3	R@5
hn	collapse	collapse	collapse
shn	54.17\pm0.05	64.99\pm0.16	69.90\pm0.21
sct	60.78\pm0.13	70.39\pm0.45	74.35\pm0.22

HOTEL50K

unzip "exp_Hotel/input/dataset.zip" for chain retrieval

Citation

@inproceedings{xuan2020hard,
  title={Hard negative examples are hard, but useful},
  author={Xuan, Hong and Stylianou, Abby and Liu, Xiaotong and Pless, Robert},
  booktitle={European Conference on Computer Vision},
  pages={126--142},
  year={2020},
  organization={Springer}
}

Updates:

Feb 14, 2021:

Improve loss function computation stablility

Improve the retrieval performance accross all datasets reported in the paper

Upload the training log information for all results

Add Hotel instance and chain retrieval code in "Hotel_instance_chain_retrieval.ipynb"

Add Hotel instance retrieval code（use in training） in "_code.Evaluation.py"