This repository provides a framework for analyzing the performance of different detectors at arbitrary frame processing rates (FPRs). We have provided sample ground truth and detections for 5 videos from the MOT Challenge Dataset, namely, Bahnhof, MOT16-02, MOT16-05, MOT16-09, MOT16-10, MOT16-11. The detections are analyzed and classified as a True/False positive by the script and a segregation is done for different pixelDistances.
For more details, have a look at our paper here. It has been accepted at WACV '18.
- Python3
- Numpy
- Matplotlib
A brief description of the files included is as follows:
| File Name | Description | Usage |
|---|---|---|
| analyseDetections.py | This script reports the True Positives/False Positives over different distances (pixelDistances) using detections from detectionResults/ and ground truth from gt/ . | Sample: ./analyseDetections.py -d detectionResults/bahnhof_ssd_mobilenet_v1_coco_11_06_out.txt -s 15 -gt gt/bahnhof_gt.txt. Use ./analyseDetections.py -h for sample usage. |
| evaluateGTfromAnnotations.py | This script reports the True Positives/False Positives over different distances (pixelDistances) using the ground truth. | Use ./evaluateGTfromAnnotations.py -h for sample usage. |
| getIDDistribution.py | This script lets you plot a histogram for an estimation of the entropy of the video | Sample: ./getIDDistribution.py -gt gt/bahnhof_gt.txt. Use ./getIDDistribution.py -h for sample usage. |