Skip to content

Latest commit

 

History

History
123 lines (95 loc) · 5.42 KB

README.md

File metadata and controls

123 lines (95 loc) · 5.42 KB
Raw

Solutions for CVWC2019 Track3&Track4 Challenge

Note:

  • Input size is 256x128
  • Multi-scale test is used.
  • Multi-model ensemble is used.
  • The training of detection is not provided, we provide the detection results.
  • Some of ReID model is pretrained on DeepFashion2. we only provide the pretrained model.
  • Arcface loss is not used.
  • We don't split the the data train and validation set.

Environment

The code is developed using python 3.7 on Ubuntu 16.04. NVIDIA GPUs are needed. The ReID models are developed and tested using 1 TITAN V GPU cards. The detection models are develop on 8xTesla V100. Other platforms or GPU cards are not fully tested.

Quick start

Our project is based on the following repositories:

Track3 briefly description:

We build a ReID model base on global features and batch hard mining triplet loss. More details, some training tricks are used, including warmup,random erasing,centerloss and etc. For post-process, we use multi-scales augmentation and reranking.

Track4 briefly description:

We build the wild ReID model by combining dectection and plain ReID model directly. The detection model is based on Cascade RCNN and coco finetune is used . The ReID model is trained with the same techique as Track3. Specially, the results of dectections is filter by the 0.05 score threshold and nms is applied. Then the retrieval process is performanced on the filtered detected images.

Installation

  1. Install pytorch >= v1.0.0 following official instruction.

  2. Clone this repo, and we'll call the directory that you cloned as ${ReID_ROOT}.

  3. Install dependencies following repo1 and repo2.

  4. Install extra dependencies:

    pip install tqdm

  5. Download pretrained models from our model zoo(GoogleDrive or OneDrive)

    ${ReID_ROOT}
 `-- model_checkpoint_pytorch
 	 `-- se_resnext50_32x4d-a260b3a4.pth
     `-- deepfashion2
         |-- Experiment4-senext50-224x224_lr2-bs8x8-warmup10-erase0_5-labelsmooth_on-laststride1-bnneck_on-triplet_center-arcface
         |   |-- se_resnext50_model_30.pth
         |   |-- se_resnext50_optimizer_30.pth
         |   |-- se_resnext50_optimizer_center_30.pth

Data preparation

**For CVWC2019 Tiger ReID Track3&4 **, please download the data to ${DATASET_ROOT_DIR} , modify data_dir in *.py and modify DATASET_ROOT_DIR in *.sh , run the following command:

python prepare_train.py
python prepare_track3_crop_train.py
python prepare_track3_crop_test.py
python prepare_track4_test.py
DATASET_ROOT_DIR=/data/nif/tiger/reid/
cp ${DATASET_ROOT_DIR}/*.csv detections/work_dirs/crop_images/htc_reidtrain_e10_2200_1100/
cd reid-strong-baseline
# train
bash sh_tiger/train.sh
# merge feat
python merge_feat.py
# track3 test
bash sh_tiger/aug_test.sh
# track4 test
bash sh_tiger/wild_aug_test.sh
bash sh_tiget/wild_sub.sh

Testing on dataset using model zoo's models(GoogleDrive or OneDrive)

download our trained model and put it in directory 'tiger/work_dirs/*/*' like this:

python prepare_train.py
# crop image a max detected area 
python prepare_track3_crop_train.py
python prepare_track3_crop_test.py
python prepare_track4_test.py
DATASET_ROOT_DIR=/data/nif/tiger/reid/
cp ${DATASET_ROOT_DIR}/*.csv detections/work_dirs/crop_images/htc_reidtrain_e10_2200_1100/
cd reid-strong-baseline

# merge feat
python merge_feat.py
# track3 test
bash sh_tiger/aug_test.sh
# track4 test
bash sh_tiger/wild_aug_test.sh
bash sh_tiget/wild_sub.sh

Ablation Study

track3

model description mAP(single_cam) mAP(cross_cam)
pcb_6 pooling vertically(softmax loss,not based the code) 0.722 0.466
senext50+alltricks(softmax + batch hard mining triplet) 0.782 0.487
senext50+alltricks+multi-scale test+ reranking 0.841 0.524
senext50+alltricks+multi-scale test+ reranking+crop image 0.835 0.519
6 ensemble model+adjust the params of reranking 0.866 0.525

track4

Model description mAP(single_cam) mAP(cross_cam)
senext50+alltricks 0.767 0.500
senext50+alltricks+multi-scale test+ reranking 0.830 0.528
senext50+alltricks+pretrain 0.777 0.479
senext50+alltricks+pretrain+multi-scale test+reranking 0.828 0.510
6 ensemble model+adjust the params of reranking 0.841 0.539