Crowd-Simulator-segmentation

Based on output generated by our branch at https://github.com/mmlab-cv/Crowd-Simulator/tree/static-simulation

Report: Computer_Vision_Project.pdf

Scripts:

• detectron_model.py : main file, does training, inference, and testing
• segMaskToCOCO.py : parses the dataset generated by Crowd Simulator (color images + segmentation masks), and creates a COCO annotation in json
• datasets/coco/bbox.py and datasets/coco/coco.py : scripts used to modify the annotations and make new filtered versions
• datasets/coco/merge.py : set of function to convert RLE annotations to polygon and to merge two annotation files.

runs folder: each run contains

• cfg.yml file containing settings used for the run
• some inference examples of the trained model (images beginning with inference_)
• evaluation_result_* for different test datasets
• metrics.json: training log
• NOTE: .pth trained models are not shared

Datasets and annotations are available from links in datasets/link_drive.txt (need unitn mail to access drive)

• train_compressed.7z contains a lossy compression (.png to .jpg) of the original synthetic dataset used in the report