Skip to content

Latest commit

 

History

History
92 lines (79 loc) · 3.14 KB

AICity2020.md

File metadata and controls

92 lines (79 loc) · 3.14 KB
Raw

Prepare Data

  • Download CityFlow and VehicleX , rename them to 'AIC20_ReID' and 'AIC20_ReID_Simulation' respectively.
  • Weakly Supervised crop Augmentation. Crop vehicle in image via weakly supervised method, a vehicle ReID pretrain model is needed to generate attention map. If you dont want to train it yourself you can get pretrain model
# first temporary comment aicity20.py line 49 # train += self._process_dir(self.train_aug_dir, self.list_train_path, self.train_label_path, relabel=False)
# to make sure only original data is used
# step 1: train inital model
python tools/train.py --config_file='configs/aicity20.yml' \
MODEL.DEVICE_ID "('2')" \
MODEL.MODEL_TYPE "baseline" \
MODEL.NAME "('resnet50_ibn_a')" \
MODEL.PRETRAIN_PATH "('/home/zxy/.cache/torch/checkpoints/resnet50_ibn_a.pth.tar')" \
SOLVER.LR_SCHEDULER 'cosine_step' \
DATALOADER.NUM_INSTANCE 16 \
MODEL.ID_LOSS_TYPE 'circle' \
SOLVER.WARMUP_ITERS 0 \
SOLVER.MAX_EPOCHS 12 \
SOLVER.COSINE_MARGIN 0.35 \
SOLVER.COSINE_SCALE 64 \
SOLVER.FREEZE_BASE_EPOCHS 2 \
MODEL.TRIPLET_LOSS_WEIGHT 1.0 \
DATASETS.TRAIN "('aicity20',)" \
DATASETS.TEST "('veri',)" \
DATASETS.ROOT_DIR "('/home/zxy/data/ReID/vehicle')" \
OUTPUT_DIR "('./output/aicity20/0326-search/augmix/')"

# step2: use inital model to crop vehicles
python tools/aicity20/weakly_supervised_crop_aug.py --config_file='configs/aicity20.yml' \
MODEL.DEVICE_ID "('0')" \
MODEL.NAME "('resnet50_ibn_a')" \
MODEL.MODEL_TYPE "baseline" \
DATASETS.TRAIN "('aicity20',)" \
DATASETS.TEST "('aicity20',)" \
DATALOADER.SAMPLER 'softmax' \
DATASETS.ROOT_DIR "('/home/zxy/data/ReID/vehicle')" \
MODEL.PRETRAIN_CHOICE "('self')" \
TEST.WEIGHT "('./output/aicity20/0326-search/augmix/best.pth')"

# AIC20_ReID_cropped will be saved at './output/aicity20/0326-search/augmix/'
# dont forget to uncomment aicity20.py line 49 # train += self._process_dir(self.train_aug_dir, self.list_train_path, self.train_label_path, relabel=False)
  • after all works have be done, data folder should look like
-Vehicle
--AIC20_ReID
--AIC20_ReID_Simulation
--AIC20_ReID_Cropped

Download pretrain model

We use ResNet-ibn as backbone. Download ImageNet pretrain model at here

Train

  • Vehicle ReID. Train three models respectively(resnet50, resnet101, resnext101),
bash ./scripts/aicity20/train.sh
  • Orientation ReID Train orientation ReID model
bash ./scripts/aicity20/ReOriID.sh
  • Camera ReID Train camera ReID model
bash ./scripts/aicity20/ReCamID.sh

you can either download our trained models

Test and ensemble

  • generate orientation and camera similarity matrix
bash ./scripts/aicity20/generate_matrix.sh
  • generate vehicle distance matrix
bash ./scripts/aicity20/test.sh
  • ensemble three distmat from three models
python ./tools/aicity20/multi_model_ensemble.py