Capstone Project

Udacity MLND

VIVA Hand Detection Challenge Using A Keras RetinaNet

Overview

In this project, RetinaNet, a one-stage object detection system developed by Lin et al is used for the VIVA Hand Detection Challenge.

The model training piggy backs off the keras-retinanet implementation.

You'll also need Octave and Piotr's Computer Vision Matlab Toolbox for evaluation.

For more detail about this project, please refer to the capstone_report.pdf.

Setup

1. Clone this repository.
2. Download and extract the VIVA Hand Detection Challenge dataset.
3. Download and extract the VIVA Hand Detection Challenge evaluation kit.
4. Create environent with the environment.yml file. (Optional)
5. In the preprocessing.ipynb update the directories to properly point to the downloaded dataset. (The test data annotations are in the evaluation kit.)
6. Download Octave or MATLAB.
7. Download Piotr's Computer Vision Matlab Toolbox.
8. Update the path in the demo.m file in the evaluation kit downloaded at step 3 to point to the toolbox downloaded above.
9. The evaluation kit file main_handdetect.m did not work for me. If you're having the same issue, try changing :

dt{currloc} = [dt{currloc};currbb];

to

[dt{currloc}] = deal([dt{currloc};currbb]);

on lines 46 and 70.

Sample Detection

VIVA Test Data

Evaluation

Some scores using the VIVA Hand Detection Challenge evaluation kit available on the challenge website.

VIVA Evaluation	L1 (AP/AR)	L2 (AP/AR)
Epoch 18	92.6/90.3	82.3/71.1
Epoch 21	91.0/90.7	78.9/73.9

The results for the snapshot at Epoch 21 ranks 4th in the leaderboard. (Ranked by L2 AR.)

Non-VIVA Challenge Detection

Not as good. Improvements can be made. (More image augmentation, larger dataset, etc)