Codebase for the paper "Detection of Systematic Errors in Object Detectors via Controlled Realistic Scene Generation."

This repository contains the codebase for our research on detecting systematic errors in object detectors through controlled realistic scene generation. Our method, BEV2EGO, leverages bird's-eye view (BEV) configurations to generate first-person view (EGO) scenes, enabling a comprehensive analysis of state-of-the-art object detection models.

Our approach allows for realistic generation of complete scenes with road-contingent control, mapping 2D BEV scene configurations to EGO perspectives. We propose a benchmark for controlled scene generation and perform a systematic analysis of multiple SOTA object detection models to identify their differences.

Code will be released soon. Instructions for setting up and running the code will be provided here.