This repository contains a PyTorch implementation of the architecture introduced in the "Real-time object detection method based on improved YOLOv4-tiny" for the final project of the Deep Learning course.
The network have been trained to detect and classify faces with and without masks. The directory datset contains images, labels and csv files used to associate them.
The directory models contains the following trained models:
model.pt, baseline model.model.tflite, TFlite conversion of the previous one.pmodel.tflite, pruned version in TFlite version.
The directory utilities contains the implementation of the loss function, functions to deal with bounding boxes, the structured pruning pipeline used and a script to compute the anchor boxes with k-means clustering algorithm.
The directory metrics contains the scripts used to compute the FPSs and MAPs.
In addition to the libraries in requirements.txt, the following libraries are required, whose installation is system-dependent:
Run
python detect/detectttf.py namemodel.tfliteto use the model converted in TFlite with the selected model or
python detect/detect.pyto use the Pytorch one and it will automatically open your webcam.
Loading the data you could experience problems with boxes' coordinates exceeding the ranges, such as:
File "/home/yourname/anaconda3/lib/python3.8/site-packagesalbumentations/augmentations/bbox_utils.py", line 328, in check_bbox
raise ValueError(
ValueError: Expected x_max for bbox (0.98828125, 0.20502645502645503, 1.0009765625, 0.2605820105820106, 1.0) to be in the range [0.0, 1.0], got 1.0009765625.)
This seems to be a regular problem with the albumentation data augmentation package. In the absence of a real solution and noticing that the error raised from very small values (~10e:-4) we have brutally get around the problem modifying the check_bbox function rounding values exceeding the margins to 0 or 1.
An additional analysis can be performed. Three different dataset are used and they must be added in the dataset_extra folder. The images of the datasets will be placed in subfolders whose full path will be:
- dataset_extra/test_set_select
- dataset_extra/test_set_batchII
- dataset_extra/fairface_dataset/train
- dataset_extra/fairface_dataset/test
Once the datasets have been downloaded and correctly stored, the analysis can be perfomed via
python extra_analysis.pyand the results will be stored in the results folder