Pathvision: An Explainable AI Method for Object Detection models

Using an object-specific tracking algorithm with state-of-the-art gradient-based Explainable AI techniques to identify classification errors in Object Detection models

Dissertation

Vanilla Gradients

Smoothgrad

Pathvision's only been developed and tested on Microsoft Windows Only

Prerequisites

• Python 3.9+
• pip 22.3.1+
• setuptools 67.7.2+
• Windows C++, C build tools
• PyTorch with CUDA support

Setup

These instructions are for Microsoft Windows 10/11 PS and cmd. (I have reasonable faith the same steps will work just fine on Unix).

Step 0: Clone the repository

Navigate to the root direction and run:

git clone https://github.com/toby-j/pathvision.git
cd pathvision

Step 1: Create venv (Optional but recommended)

I reccomend using a venv just for simplisity.

python -m venv venv
venv/Scripts/activate
pip install .

This will run the setup.py file which'll install all of Pathvision's dependencies as specified in requirements.txt.

Step 2: Clone Detectron2

cd pathvision/models
git clone https://github.com/facebookresearch/detectron2.git detectron2

Step 3: Build Detectron2

This repository is using Detectron2 for image segmentation, which needs to be built on your machine. Detectron2 requires CUDA to be built. You check check that your installed PyTorch has CUDA enabled with:

python -c 'import torch; from torch.utils.cpp_extension import CUDA_HOME; print(torch.cuda.is_available(), CUDA_HOME)'

If this prints True you'll be able to build Detectron2 If it prints false, you may need to reinstall PyTorch with CUDA compute platform enabled, see: Install PyTorch locally.

Navigate to the Detectron2 folder:

cd pathvision/models/detectron2
pip install .

We've now built and installed the dependencies for Pathvision.

Unfortunately, I can't find a way to build Detectron2 without CUDA. On Linux you can use the pre-builds, see: Detectron2 Docs. You then place the files inside of pathvision/models/detectron2

Running

There's a tester.py file which creates an instance of Pathvision's Object Detection feature and runs a predefined test using images from debug/test_data. If you want to use your own test images, simply add a folder in debug/test_data and change the folder path of frame_list_dir in tester.py to the folder containing your test images (.jpg and .png are both supported).

You can select either VanillaGradients, IntegratedGradients or Smoothgrad. If you want fast results, usually a minute or so per frame, use vanilla gradients. Smoothgrad is usually a couple of minutes. Integrated Gradients is much more computationaly expensive, but is supported.

To view results, the output images will be written to debug/final_output and debug/output_image. The tracking logs and gradient overlap calculations will be written in the debug/log folder.