Berkeley DeepDrive Image Segmentation Attempt

Work in Progress

This was my attempt to create an image segmentation model using Berkeley's DeepDrive dataset. A more complete writeup documenting the journey can be found in my medium post. https://medium.com/p/308f8c44305a/edit

The original dataset can be downloaded here: https://bdd-data.berkeley.edu/

Installation

To recreate my results you'll need your Linux distro of choice, PyTorch v1 and Python 3.6 or later.

$ conda install -c pytorch -c fastai fastai

From there Berkeley DeepDrive v2.ipynb should run. v1 was created using an earlier version of FastAI and did not successfully segment.

Included

label_quantify.py

• Was used to determine how many categories there were. Companion to test_label_quantify.json

seg_128 folder

• A 128x128 bordered version of the segmentation dataset used to train ResNet34. Could be used as a quick reference to try deeper ResNet or other pretrained models.

Task list

• Implement U-Net
• Implement 100-layer Tiramisu
• Mask R-CNN
  • Note: Explored but no implementation attempted