[target detection] | indicator a probabilistic challenge for object detection

A probabilistic challenge for object detection

Target detection failure can distinguish four main types of object detection failure ： Failed to detect object ; Assign the wrong category label to the detected object ; Positioning objects incorrectly - for example , Only some objects are detected ; And incorrectly detecting nonexistent objects .

In order to run safely and reliably in robots or autonomous systems ,CNN Must express meaningful Uncertain information . The target detector must quantify the report Uncertainty of labels and bounding boxes . However , Although state-of-the-art object detectors report at least one uncalibrated label uncertainty indicator by label distribution or label score , But they currently do not report spatial uncertainty . Assessing label quality or spatial uncertainty is not even within the scope of typical computer vision benchmarking and competition . Put forward an evaluation standard and data set .
（Probabilistic Object Detection: Definition and Evaluation
The Probabilistic Object Detection Challenge
）

As the evaluation results show , When combined with existing target detectors , Monte Carlo attenuation and other uncertainty techniques show good prospects , But we should develop mature probabilistic target detection methods , A lot of research is needed .