SeaBird: Segmentation in Bird's View with Dice Loss Improves Monocular 3D Detection of Large Objects

Monocular 3D detectors achieve remarkable performance on cars and smallerobjects. However, their performance drops on larger objects, leading to fatalaccidents. Some attribute the failures to training data scarcity or theirreceptive field requirements of large objects. In this paper, we highlight thisunderstudied problem of generalization to large objects. We find that modernfrontal detectors struggle to generalize to large objects even on nearlybalanced datasets. We argue that the cause of failure is the sensitivity ofdepth regression losses to noise of larger objects. To bridge this gap, wecomprehensively investigate regression and dice losses, examining theirrobustness under varying error levels and object sizes. We mathematically provethat the dice loss leads to superior noise-robustness and model convergence forlarge objects compared to regression losses for a simplified case. Leveragingour theoretical insights, we propose SeaBird (Segmentation in Bird's View) asthe first step towards generalizing to large objects. SeaBird effectivelyintegrates BEV segmentation on foreground objects for 3D detection, with thesegmentation head trained with the dice loss. SeaBird achieves SoTA results onthe KITTI-360 leaderboard and improves existing detectors on the nuScenesleaderboard, particularly for large objects. Code and models athttps://github.com/abhi1kumar/SeaBird