Localization Uncertainty Estimation for Anchor-Free Object Detection

Since many safety-critical systems, such as surgical robots and autonomous driving cars, are in unstable environments with sensor noise and incomplete data, it is desirable for object detectors to take into account the confidence of localization prediction. There are three limitations of the prior uncertainty estimation methods for anchor-based object detection. 1) They model the uncertainty based on object properties having different characteristics, such as location (center point) and scale (width, height). 2) they model a box offset and ground-truth as Gaussian distribution and Dirac delta distribution, which leads to the model misspecification problem. Because the Dirac delta distribution is not exactly represented as Gaussian, i.e., for any $\mu$ and $\Sigma$. 3) Since anchor-based methods are sensitive to hyper-parameters of anchor, the localization uncertainty modeling is also sensitive to these parameters. Therefore, we propose a new localization uncertainty estimation method called Gaussian-FCOS for anchor-free object detection. Our method captures the uncertainty based on four directions of box offsets~(left, right, top, bottom) that have similar properties, which enables to capture which direction is uncertain and provide a quantitative value in range~[0, 1]. To this end, we design a new uncertainty loss, negative power log-likelihood loss, to measure uncertainty by weighting IoU to the likelihood loss, which alleviates the model misspecification problem. Experiments on COCO datasets demonstrate that our Gaussian-FCOS reduces false positives and finds more missing-objects by mitigating over-confidence scores with the estimated uncertainty. We hope Gaussian-FCOS serves as a crucial component for the reliability-required task.