Localization Uncertainty Estimation for Anchor-Free Object Detection

Since many safety-critical systems, such as surgical robots and autonomous driving cars operate in unstable environments with sensor noise incomplete data, it is desirable for object detectors to take the localization uncertainty into account. However, there are several limitations of existing estimation methods anchor-based detection. 1) They model heterogeneous properties different characteristics scales, location (center point) scale (width, height), which could be difficult estimate. 2) box offsets Gaussian distributions, not compatible ground truth bounding boxes that follow Dirac delta distribution. 3) sensitive anchor hyper-parameters, their also highly choice hyper-parameters. To tackle these limitations, we propose a new method called UAD anchor-free Our captures four directions (left, right, top, bottom) homogeneous, so can tell direction uncertain, provide quantitative value [0, 1]. enable estimation, design loss, negative power log-likelihood measure by weighting likelihood loss its IoU, alleviates misspecification problem. Furthermore, an uncertainty-aware focal reflecting estimated classification score. Experimental results on COCO datasets demonstrate our significantly improves FCOS [32], up 1.8 points, without sacrificing computational efficiency. We hope proposed serve crucial component detection tasks.