howl0893 / custom-object-detection-datasets

Creating custom object detection datasets in the formats of COCO and KITTI

Creating Custom Object Detection Datasets

1. Create custom COCO datasets from scratch

https://www.youtube.com/watch?v=jftZBfMZj8k&list=PLy-L-4xQRIuNm4tN9RIufCC6sqjuw1mAj

2. Convert COCO dataset into KITTI format

https://www.youtube.com/watch?v=N6D9W-ytQc4&feature=youtu.be

3. Convert PASCAL VOC formatted datasets into a COCO dataset

https://www.youtube.com/watch?v=NPnPn-buQRM&list=PLy-L-4xQRIuOZFQ3UMFTDp8m_7ekbNiuP

custom-object-detection-datasets's Issues

onyl one boundry box, instead of multiple

Hi, thank you for the tutorials!
I've copied your written guide for creating a coco dataset and run into following problem:
When I have an image with multiple instances of the same object (in my case fishes), the coco file, does not have multiple boundary boxes , but only one large box, containing all found fishes. The masks with the precise shapes work fine.

In gimp i added for every mask a new layer, however used the same color for every fish, asloong as they were not intersecting. For export i checked the "save layer offset" (i first had it not checked, and then tried it as a first attempt to solve the probem). Only thing that works was when i am using different colors for every single fish mask.
Any idea?

157 line bug

To fix the bug, we changed "./python/coco_json_utils.py", line 157 to
contours = measure.find_contours(np.array(mask), 0.5, positive_orientation='low')

custom image 3d annotation to 3d kitti format

I've annotated images using 3D BOUNDING BOXES. How can I convert to 3D KITTI DATASET format?