Reimplementation of YOLOv3 with quadrangle
This is a reimplementation of YOLOv3: An Incremental Improvement and is based on Ultralytics LLC's PyTorch implementation. This work detects obejcts in arbitrary directions with quadrangle, and implemented on ICDAR2015 text dataset for example.
- Python3
- numpy
- torch
- opencv-python
- Shapely
Check the -data_config_path and -cfg in train.py.
Dataset folder is organized as follows:
- Dataset
- images
- labels
The label format: class x1 y1 x2 y2 x3 y3 x4 y4
For example (0 493 115 519 115 519 131 493 131)
$ python3 train.py
- Modify
yolov3.cfg
file- Change
[yolo]
classes with the number of classes in your own dataset. - Replace the value of filters in
[convolutional]
which lays above[yolo]
, filters should be3 * (8 + 1 + num_classes)
, where8
means 8 offsets of the quadrangle,1
means objectness confidence.
- Change
- Modify
cfg/*.data
where classes field should be your number of classes in your dataset - Modify
data/*names
and put your labels in it.
Checkpoints are saved in weights.
$ python3 detect.py