TLfPS

Transfer Learning for Person Search

Installation

• tensorflow 1.14.0
• keras
• scikit-learn [conda install -c anaconda scikit-learn]
• cuml[conda install -c rapidsai cuml]
• wxPython [conda install -c anaconda wxpython]
• yolov3_reid.h5
• mrcnn.h5
• yolov3.h5
• yolov4.h5
• yolov4_reid.h5
• mrcnn_reid.h5

Simple application

Download yolov3_reid.h5, save to ./saved_weights

#model: [yolov3], gpu: [0]
#ctrl+L: load model; ctrl+Q: select query image & draw roi; ctrl+G:select gallery images; ctrl+S:search
python app/app.py -m yolov3 -g 0

Evaluation

• Download yolov3_reid.h5, save to ./saved_weights
• Download CUHK_SYSU dataset
• Download PRW dataset

evaluate searching score for PRW and CUHK-SYSU

#model: [yolov3], gpu: [0], dataset: [path/to/CUHK_SYSU], experiment name: [cuhk_default]; evaluate on CUHK-SYSU dataset
python utils/evaluation.py -m yolov3 -g 0 -p path/to/CUHK_SYSU -e cuhk_default
#model: [yolov3], gpu: [0], dataset: [path/to/PRW], experiment name: [default]; evaluate on PRW dataset
python utils/prw_evaluation.py -m yolov3 -g 0 -p path/to/PRW -e prw_default

visualization for CUHK-SYSU

#model: [yolov3], gpu: [0], dataset: [path/to/CUHK_SYSU], experiment name: [cuhk_default]; TSNE visualization for CUHK-SYSU dataset
python utils/tsne_cuhk.py -m yolov3 -g 0 -p path/to/CUHK_SYSU -e cuhk_default
python utils/TSNE.py -m yolov3 -g 0 -p cuhk -e cuhk_default
python utils/failure_cases.py -m yolov3 -g 0 -c path/to/CUHK_SYSU -e cuhk_default

• 48 failure cases in CUHK-SYSU. (3 images a group, first is query, then false positive, then true negative. note there are some mistakes from the dataset)

visualization for PRW

#model: [yolov3], gpu: [0], dataset: [path/to/PRW], experiment name: [prw_default]; TSNE visualization for PRW dataset
python utils/tsne_prw.py -m yolov3 -g 0 -p path/to/PRW -e prw_default
python utils/TSNE.py -m yolov3 -g 0 -p prw -e prw_default
python utils/failure_cases.py -m yolov3 -g 0 -p path/to/PRW -e prw_default

• TSNE visualization for PRW, point version
• TSNE visualization for PRW, image version, high resolution image is able to download
• 48 failure cases in PRW. (3 images a group, first is query, then false positive, then true negative. note there are some mistakes from the dataset)

evaluate running speed

#model: [yolov3], gpu: [0], dataset: [path/to/CUHK_SYSU], reid: [0]; test the running time without reid module
python utils/elapsedtime.py -m yolov3 -g 0 -p [path/to/CUHK_SYSU] -r 0
#model: [yolov3], gpu: [0], dataset: [path/to/CUHK_SYSU], reid: [1]; test the running time with reid module
python utils/elapsedtime.py -m yolov3 -g 0 -p [path/to/CUHK_SYSU] -r 1

Training

• Download yolov3_reid.h5, save to ./saved_weights
• Download CUHK_SYSU dataset
• Download PRW dataset

#model: [yolov3], gpu: [0]; convert yolov3_reid.h5 to yolov3.h5
python utils/getmodel.py -m yolov3 -g 0
#model: [yolov3], gpu: [0]; test yolov3 detection network
python utils/testmodel.py -m yolov3 -g 0 -p path/to/image
#model: [yolov3], gpu: [0], dataset: [path/to/CUHK_SYSU], [path/to/PRW]; create detection, help to evaluate the running time evaluation
python utils/creategallery.py -m yolov3 -g 0 -c path/to/CUHK_SYSU -p path/to/PRW
#model: [yolov3], gpu: [0], dataset: [path/to/CUHK_SYSU], [path/to/PRW], model name: [1]; start training
python train.py -m yolov3 -g 0 -c path/to/CUHK_SYSU -p path/to/PRW -n 1
#model: [yolov3], gpu: [0], weights path: [path/to/yolov3_model-1-x-0.xxx.h5]; convert a trained model yolov3_xxx.h5 to yolov3_reid.h5
python utils/convertor.py -m yolov3 -g 0 -p path/to/yolov3_model-1-x-0.xxx.h5
'''
if you meet the [none gradient] problem, please modify the functon keras->optimizer.py->get_gradients by this:
def get_gradients(self, loss, params):
    grads = K.gradients(loss, params)
    grads = [g if g is not None else tf.constant(0, dtype = 'float32') for g in grads]
'''

Training for YOLOv4, Mask RCNN, DLA-34 etc.

• get weights from the corresponding projects and convert them to .h5 format
• put .h5 weights in pretrianed_weights with name [yolov4.h5], [mrcnn.h5], [dla_34.h5]
• change [yolov3] to [yolov4], [mrcnn], [dla-34] etc. The training process is the same with the training of [yolov3]
• for convenience， we provide the well-trained weights, yolov4_reid.h5, dla_34_reid.h5

Train YOLOv3 detection network