Berkeley DeepDrive Practice

Raymond-Wu · December 29, 2018, 4:06am

Hello all, to practice my skills with part 2 I wanted to attempt a U-Net on Berkeley’s DeepDrive dataset. However, I am getting very bad results with a mask accuracy of 13%! One of the concerns I had was that when I tried to overlay the training images and masks they weren’t matching. Could that be why the accuracy is so low?

Link to notebook:

github.com

RaymondDashWu/Berkeley-DeepDrive/blob/master/Berkeley DeepDrive.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Misc Notes: \n",
    "- All images for color_labels, images, labels 1280x720\n",
    "- Currently trying 128 resize with padding to 128x128\n",
    "    - Initially had padding as white but then realized that could conflict with labels so changed all to skyblue\n",
    "\n",
    "### Todo:\n",
    "- Images not lining up with masks?\n",
    "    - trn/trn_mask are in order so model is likely to be training in correct order\n",
    "- Possibly try resnet50"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,

This file has been truncated. show original

Link to data:

kcturgutlu · January 8, 2019, 7:30pm

It’s almost always necessary to do pre and post image plotting for sanity checks while working with vision problems. If your ground truth is somehow incorrect you would get bad segmentations in my opinion.

By looking at your notebook I assume you are using fastai v 0.7, I would suggest you to use v1.0.

Here is the notebook for camvid (SOTA results), an almost identical task to yours.

github.com

fastai/course-v3/blob/master/nbs/dl1/lesson3-camvid.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Image segmentation with CamVid"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "%reload_ext autoreload\n",
    "%autoreload 2\n",
    "%matplotlib inline"
   ]
  },

This file has been truncated. show original

Raymond-Wu · January 8, 2019, 8:46pm

Oh wow I didn’t know this existed! This looks very promising! Do you know if there’s a video associated with this notebook? I’ve been trying to implement a Mask R-CNN but I’m excited to know that there’s a FastAI notebook I can work off of.

I will take your recommendations as well regarding regarding 1.0. Thanks again for pointing me in the right direction!