xrosliang / open-reid-tracking

open-reid with PCB, IDE, triplet, ZJU; MOT/MTMCT feature extraction support included

Open-ReID-tracking

This repo is based on Cysu's open-reid, which is a great re-ID library. For performance, we implemented some other baseline models on top of it. For utility, we add some function for the tracking-by-detection workflow in tracking works.

• update all models for performance & readability.
• add data/README.md. check for folder structure & dataset download.
• add requirements.txt. use conda install --file requirements.txt to install.
• add BN after feature layer in reid/models/IDE_model.py for separation. This introduces a higher performance.
• fix high cpu usage via adding os.environ['OMP_NUM_THREADS'] = '1' in runable files.
• NEW: We adopt a baseline from Hao Luo [git, paper]. See ZJU.py. We achieve competitive performance with the same IDE_model.py.

Please use this repo alongside with our flavor of DeepCC tracker for tracking.

Model

• IDE [paper]
• Triplet [paper]
• PCB [git, paper]
• ZJU [git, paper]

Data

The re-ID datasets should be stored in a file structure like this:

~
└───Data
    └───AIC19
    │   │ track-1 data
    │   │ ...
    │
    └───AIC19-reid
    │   │ track-2 data
    │   │ ...
    │
    └───VeRi
    │   │ ...
    │
    └───DukeMTMC-reID
    │   │ ...
    │
    └───Market-1501-v15.09.15
        │ ...

Usage

Re-ID

training from scratch

CUDA_VISIBLE_DEVICES=0 python3 IDE.py -d market1501 --train

this will automatically save your logs at ./logs/ide/market1501/YYYY-MM-DD_HH-MM-SS, where YYYY-MM-DD_HH-MM-SS is the time stamp when the training started.

resume & evaluate

CUDA_VISIBLE_DEVICES=0 python3 IDE.py -d market1501 --resume YYYY-MM-DD_HH-MM-SS

Feature Extraction for Tracking (to be updated)

We describe the workflow for a simple model. For the full ensemble model, please check

First, please use the following to extract detection bounding boxes from videos.

python3 reid/prepare/extract_bbox.py

Next, train the baseline on re-ID data from AI-City 2019 (track-2).

# train
CUDA_VISIBLE_DEVICES=0,1 python3 ZJU.py --train -d aic_reid --logs-dir logs/ZJU/256/aic_reid/lr001_colorjitter --colorjitter  --height 256 --width 256 --lr 0.01 --step-size 30,60,80 --warmup 10 --LSR --backbone densenet121 --features 256 --BNneck -s 1 -b 64 --epochs 120

Then, the detection bounding box feature are computed.

# gt feat (optional)
# CUDA_VISIBLE_DEVICES=0,1 python3 save_cnn_feature.py -a zju --backbone densenet121 --resume logs/ZJU/256/aic_reid/lr001_colorjitter/model_best.pth.tar --features 256 --height 256 --width 256 --l0_name zju_lr001_colorjitter_256 --BNneck -s 1 -d aic --type gt_all -b 64
# reid feat (parameter tuning, see DeepCC_aic)
CUDA_VISIBLE_DEVICES=0,1 python3 save_cnn_feature.py -a zju --backbone densenet121 --resume logs/ZJU/256/aic_reid/lr001_colorjitter/model_best.pth.tar --features 256 --height 256 --width 256 --l0_name zju_lr001_colorjitter_256 --BNneck -s 1 -d aic --type gt_mini -b 64
# det feat (tracking pre-requisite, see DeepCC_aic)
CUDA_VISIBLE_DEVICES=0,1 python3 save_cnn_feature.py -a zju --backbone densenet121 --resume logs/ZJU/256/aic_reid/lr001_colorjitter/model_best.pth.tar --features 256 --height 256 --width 256 --l0_name zju_lr001_colorjitter_256 --BNneck -s 1 -d aic --type detections --det_time trainval -b 64
CUDA_VISIBLE_DEVICES=0,1 python3 save_cnn_feature.py -a zju --backbone densenet121 --resume logs/ZJU/256/aic_reid/lr001_colorjitter/model_best.pth.tar --features 256 --height 256 --width 256 --l0_name zju_lr001_colorjitter_256 --BNneck -s 1 -d aic --type detections --det_time test -b 64

Implementation details

Cross-entropy loss:

• batch_size = 64.
• learning rate = 0.1, step decay after 40 epochs. Train for 60 epochs in total.
• 0.1x learning rate for resnet-50 base.
• weight decay = 5e-4.
• SGD optimizer, momentum = 0.9, nestrov = true.

Triplet loss:

• margin=0.3.
• ims_per_id = 4, ids_per_batch = 32.
• learning rate = 2e-4, exponentially decay after 150 epochs. Train for 300 epochs in total.
• unifide learning rate for resnet-50 base and fc feature layer.
• weight decay = 5e-4.
• Adam optimizer.

Default Settings:

• IDE
  • stride = 2 in last conv block.
  • h x w = 256 x 128.
  • random horizontal flip + random crop.
• Triplet
  • stride = 2 in last conv block.
  • h x w = 256 x 128.
  • random horizontal flip + random crop.
• PCB
  • stride = 1 in last conv block.
  • h x w = 384 x 128.
  • random horizontal flip.
• ZJU
  • cross entropy + triplet.
  • ims_per_id = 4, ids_per_batch = 16.
  • h x w = 256 x 128.
  • warmup for 10 epochs.
  • random horizontal flip + pad 10 pixel then random crop + random erasing with re = 0.5.
  • label smooth.
  • stride = 1 in last conv block.
  • BNneck.
  • center loss.

Tracking settings for IDE, Triplet, and PCB:

• stride = 1 in last conv block.
• h x w = 384 x 128.
• horizontal flipping + Random Erasing with re = 0.5.

Raw setting for ZJU:

• cross entropy + triplet.
• ims_per_id = 4, ids_per_batch = 16.
• h x w = 256 x 128.
• random horizontal flip + pad 10 pixel then random crop.

Experiment Results

dataset	model	settings	mAP (%)	Rank-1 (%)
Duke	IDE	Default	58.70	77.56
Duke	Triplet	Default	62.40	78.19
Duke	PCB	Default	68.72	83.12
Duke	ZJU	Default	75.20	86.71
Market	IDE	Default	69.34	86.58
Market	Triplet	Default	72.42	86.55
Market	PCB	Default	77.53	92.52
Market	ZJU	Default	85.37	93.79