An app which solves problem of object localization of a ball on videos of me shooting on basket. Results are given below:
Problem is solved using deep neural net
with a backbone ConvNeXt on which was added a couple of dense layers as Head with the sigmoid as final activation.
The output od model are 4 number representing x1, y1, x2 and y2 coordinates. Backbone was not pretrained, weights were randomly initialized
coming from normal distribution. Training was done on one GPU and lasted around 1-2 hours. Dataset was annotated by me.
My goal was to also implement hyperas and autokeras libraries in my model to perform NAS and some HPO, but I did not succeed in that. Those libraries seemed incompatible with newer architectures, at least thats how I interpreted it.
I used mIOU as metric, and as it can be seen on the results I didnt get perfect results. My biggest test mIOU was around 0.65, but I am pretty satisfied with the results.
Clone the repo:
https://github.com/babotrojka/BasketPredictor.git
Install reqs preferably in a conda env:
pip install -r requirements.txt
Create and structure dataset like this:
dataset
|___train
| | frame_0001.jpg
| | frame_0001.json
| | ...
|
|___val
|
|___test
JSON file looks like this:
"bbox": [0.4287037037037037, 0.19166666666666668, 0.4962962962962963, 0.23333333333333334]}
Dataset folder should be situated in the parent folder of the repo or in train.py DATASET_ROOT could be updated to a new dataset root.
Training can then be started with:
python train.py
My idea was also to implement video classification in the classes marking if the basket was scored or not. The idea was to use Transformer architecture. Unfortunately, other projects and obligations came along...