In this project we focused on manhole detection for self-driving vehicles. The dataset collected from various sources is minimal, with only 1032 images and annotations.
-
CNN for classification
-
Visual Transformers (ViT) for classification
-
Classical CV for localization followed by ANN for classification
-
YOLOv3 and YOLOv3-tiny for object detection (CNN backbone)
-
YOLOS for object detection (Transformer backbone)
There is also a script for detection of size of the hole in a broken/open manhole with respect to the size of the manhole. If the size of the hole is greater than 50% of the manhole size, the hole is recorded as an important one in need of fixing.
CONTRIBUTORS:
- Shika Rao
- Nitya Mitnala