YOLOv3 object detection and explanation with RISE and GradCAM
This repo has implementation of YOLOv3 in Keras. The code is inspired by below implementations
- https://github.com/david8862/keras-YOLOv3-model-set
- https://github.com/experiencor/keras-yolo3
The software/tools I have used
- python 3.6
- tensorflow 2.3.1
- keras 2.4.3
- tensorboard 2.3.0
I haven't directly forked these repos rather borrowed few function and modules to aujust my implementation. This is fully custom trained YOLOv3 model with explanation in Grad-CAM and RISE. The ipyhon note for Grad-CAM and RISE is available here
Since, this repo is mainly my masters dissertation from LUMU, there are lots of testing code which I have used to in Atom editor with hydrogen kernel enabled.
The structure of the repo I kept like below
jmod/
├── keypoint
│ ├── centernet
│ │ ├── augmentor
│ │ │ └── __pycache__
│ │ ├── eval
│ │ ├── generators
│ │ │ └── __pycache__
│ │ ├── models
│ │ │ └── __pycache__
│ │ ├── __pycache__
│ │ └── utils
│ │ └── __pycache__
│ └── __pycache__
├── onestage
│ ├── __pycache__
│ └── yolov3
│ ├── models
│ │ └── __pycache__
│ ├── __pycache__
│ └── result
│ └── classes
└── __pycache__
Under YOLOv3 I have the complete implementation. The main training module/file is
The inferensr or evaluation module/file is
This implementation is for MIO-TCD dataset available here
Since the entire dataset is huge, I have serialized the train set after preprocessing into dictionary format which is available below
- train set instances: https://github.com/jmajumde/MyMScProj/tree/main/dataset_serialized
The dataset can be loaded with below sample python code
with open(os.path.join(proj_dir_path,"all_insts"), "rb") as rh:
train_ds = pickle.load(rh)
rh.close()
with open(os.path.join(proj_dir_path,"seen_labels"), "rb") as rh:
seen_train_labels = pickle.load(rh)
rh.close()The train_ds data dictionary then looks like below
train_ds[:2]
[
{'object': [{'name': 'car', 'xmin': 405, 'ymin': 79, 'xmax': 565, 'ymax': 155}, {'name': 'car', 'xmin': 266, 'ymin': 160, 'xmax': 528, 'ymax': 271}], 'filename': '/mywork/PGDDS-IIITB/MyDatasets/MIO-TCD/MIO-TCD-Localization/train/00101947.jpg', 'height': 480, 'width': 720},
{'object': [{'name': 'pickup_truck', 'xmin': 214, 'ymin': 323, 'xmax': 441, 'ymax': 458}, {'name': 'single_unit_truck', 'xmin': 57, 'ymin': 132, 'xmax': 127, 'ymax': 207}], 'filename': '/mywork/PGDDS-IIITB/MyDatasets/MIO-TCD/MIO-TCD-Localization/train/00080403.jpg', 'height': 480, 'width': 720}
]
The model is working but there are still pending work to make the evaluation works