Object Tracking
- Simultaneous Multiple Object Detection and Tracking System in Keras (Detection network based on YOLOv2 - reimplemented in keras)
- Single Object Tracking with FasterRCNN and YOLOv2/v3 as detection backends
Trackers Available:
- TinyTracker (Static Detection Priors from FasterRCNN or Yolov2/v3)[Single Object]:
- TinyHeatmapTracker (Static Detection Priors from FasterRCNN or Yolov2/v3)[Single Object]:
- MultiObjDetTracker (Trainable Detection Priors from Yolov2 reimplmented in Keras)[Multiple Objects]:
NOTE: Yolov2 reimplementation in Keras as standalone detector also available
Dependencies
- Tensorflow
- Keras
- OpenCV
- easydict (for py-faster-rcnn)
- cython (for py-faster-rcnn)
- imgaug
Instllation
-
Run git clone --recursive https://github.com/kshitiz38/object-tracking.git
- NOTE: If you didn't clone with the --recursive flag run manually the following code
git submodule update --init --recursive
-
Darknet
-
Faster RCNN
Usage
- For Single Object Tracking
- Modify Parameters in config.jon
- Convert Datasets to PASCAL VOC format if not already
- Run
python utility/tb_to_pascal.py' or 'python utility/tb_to_pascal.py
or write one for your own dataset
- Run
python trainer.py
- For Simultaneous Multiple Object Detection and Tracking
- Modify Parameters in
KerasYOLO.py
and MultiObjDetTracker.py
- Convert Datasets like above specify paths in
MultiObjDetTracker.py
already done for ImageNet Vid and MOT17
- Run
python trainer.py
NOTE :
- Call
single_object_tracking()
in trainer.py
for Single Object Detection with fixed detection priors from Other Detection backends
- Call
simult_multi_obj_detection_tracking()
in trainer.py
for Simultaneous Multiple Object Detction and Tracking with Yolov2 Reimplemented in Keras
Model Architectures
- Coming Soon!!
- Feel free to figure out yourself!! See
models_tracking
and models_detection
directories
TODOs
- [ ] Add theory and model architectures explaination
- [ ] Add config.json file for parameters for MultiObjDetTracker and KerasYOLO
- [ ] Benchmark for ImagenetVid Challenge, MOT and VisualTB Datasets
- [ ] Add support for Detectron models as detection backend
References
- https://github.com/Guanghan/ROLO
- https://github.com/experiencor/keras-yolo2