Accurate object detection is fundamental to the safety and reliability of autonomous driving systems. Among anchor-free frameworks, FCOS has proven to be an effective baseline thanks to its simplicity and dense per-pixel prediction strategy. In FCOS, bounding-box regression estimates the offsets from each reference point to the four object sides, typically supervised by the Intersection over Union (IoU) loss, which jointly penalizes all four offsets. However, these offsets are not fully correlated, as each corresponds to a distinct region of the object. Moreover, IoU loss combines object size and shape into a single penalty term, making it difficult for the model to distinguish whether localization errors arise from scale mismatch or shape distortion. To address these limitations, we introduce Polar Loss , a novel IoU-free regression function that reformulates bounding-box prediction by estimating the positions of two diagonally opposite corners. This formulation explicitly decouples the learning of opposing offsets, allowing each side of the box to be optimized independently. Furthermore, Polar Loss expresses each corner in polar coordinates, where the radial distance represents object scale and the angular component captures shape variation. This enables the model to learn size and shape through distinct and interpretable terms. When integrated into the FCOS detector, Polar Loss improves accuracy across two autonomous-driving benchmarks, achieving +0.5% Average Precision on BDD100K and +1.1% on TJU-DHD without increasing inference cost. These gains reflect more accurate bounding-box regression, leading to enhanced perception reliability, which is particularly critical for safety in autonomous driving.
Mboutayeb et al. (Fri,) studied this question.