idd_driving_detection
Indian Driving Dataset Road Objects Detection
workflow
- Convert the images and annotations names to '0.xml/0.jpg', '1.xml/1.jpg' and so on.
- Use xml_csv.py file to convert the xml format to csv/txt format.
Codes to work upon
Look for repo containing cityscapes dataset and run the codes anyhow Cityscapes have void classes look for it
- divamgupta github repo -> Try segnet and all rest of the models
- qubvel segmentation models
- Efficient Net
- PSP Net
- Deep Lab V3
- kNN -> Nearest pixel -> Class identification + Dealing with void pixels
Problems Currently facing
- How to deal with unlabeled pixels in image segmentation?
- Dealing with instance labels in the dataset
Concepts to look upon
- FCN-8, 16, 32 difference
1st set of results Validation:
Level 1 Scores | mIoU |
---|---|
drivable | 0.7845120106595769 |
non-drivable | 0.23014790726948495 |
living-things | 0.3588385523008605 |
vehicles | 0.6218017694882831 |
road-side-objs | 0.34466576972739665 |
far-objects | 0.6662080565682629 |
sky | 0.8938274671391853 |
mIoU | 0.4875001916441313 |
[0.78466189 0.23014791 0.35885548 0.6218409 0.34473394 0.66637605 0.89399533]
mIoU: 55.723021395073424
Submissions
S. No. | Architecture | mIoU | Comments |
---|---|---|---|
1 | Res50+UNet | 0.404 | 10 epochs + residual pixels==0 |
Solved
- Change the intensity of particular values
- Change the name to label.png
- Directories make
- Test folder some images are not present -> 5 images ID find out