Metric calculator for oriented bounding boxes and object occupancy accuracy.

Introduction

This toolkit calculates several metrics related to object 3D pose estimation and its corresponding occupancy accuracy.

For oriented bounding boxes:

• AP_3D (Iou = 0.25, 0.5, etc.), average precision for 3D detection.
• AOS (Iou = 0.25, 0.5, etc.), average orientation similarity.
• OS, orientation ocore.

For object occupancy:

• MODA, multi-object detection accuracy.
• MODP, multi-object detection precision.
• Precision.
• Recall.

Requirements

• Install module shapely using pip install shapely.
• Install MATLAB and MATLAB engine, link.

Data Preparation and Usage

Both of your ground truth data and results should be capsuled in text file and formatted into a n,14 matrix as below, where n represents the number of samples you have in you data:

frame num | occ_x | occ_y | w | h | r_x1 | r_y1 | r_x2 | r_y2 | r_x3 | r_y3 | r_x4 | r_y4 | degree

where

• frame num refers to the frame identity number,
• occ_x & occ_y refers to object global coordinate on occupacy map,
• w & h refers to object origin width and height,
• r_x/r_y refer to the rotated coordinates of the bounding boxes.
• degree refers to the orientation degree of the current bounding box.

Please refer to example gt.txt and results.txt for sample data, evaluate.py to specify your file path.

Run the script using cd code && evaluate.py to evaluate the template results, the expected output should be:

Parameter settings

Specify distance threshold (in centimeters) for calculating MODA and MODP in CLEAR_MOD_HUN.m line 35.