Assignment for Object Detection during VR course at MIMUW in 2022, graded for 10 / 10
Beware as repo is ~400MB large, as it consists of two fully trained models
Author: Marcin Mazur
models- best model checkpoints.plots- plots generated during trainings.report- report with techinques, summary of my work (mm418420.pdf).srcsource files.data- module for data processing.utils- module for useful help functions.
efficient_det.ipynb- notebook with logs from best run.
Tests should be fast ~1min, and can be run with pytest command.
Full training (10 epochs) takes ~3min on 4GB GPU.
Project was developed on:
- python 3.9.7
- pytorch 1.11.0
In this task - one should implement a part of the one of the state-of-the-art NN architecture for dense predictions - EfficientDet - called the BiFPN feature combinator. Then one should use the features from the FPN as an input to two heads: one for an person segmentation (single class) and other for an edge detection. This network should be trained on the PennFundan dataset with traditional augmentations.