We are thrilled to be presenting our #CVPR2024 paper today! Stop by our poster - we're eager to share our findings and hear your thoughts!.🗓 Thursday 20th June 2024.🕥 10:30 AM to 12:00 noon (Pacific Time).📍 Exhibit Hall (Room 4A-E), Seattle Arch Convention Center.🏁 58.

📢The @neur_reps Workshop @NeurIPSConf 2024 extends submission deadline. Consider submitting your research on incorporating physical laws, symmetry, geometry, or topology in neural nets to this exciting venue.

1page Resume

2Page Resume:.

I'm officially on the industry job market! I am looking for full time Research Scientist,  Research Engineer and Applied Scientist positions starting January 2025. I like working on 3D problems like 3D detection and 3D reconstruction. If you find my profile interesting, DM!.

particularly for large objects (> 8m in length). 13/n

and improves existing detectors on the nuScenes leaderboard. 12/n

SeaBird achieves SoTA results on the KITTI-360 leaderboard. 11/n

Leveraging our theoretical insights, we propose SeaBird (Segmentation in Bird's View) as a first step towards generalizing to large objects. SeaBird effectively integrates BEV segmentation on foreground objects for 3D detection with segmentation head trained with dice loss. 10/n

We mathematically prove that the dice loss leads to superior noise-robustness and model convergence for large objects compared to regression losses for a simplified case. 9/n

We argue that the cause of failure is the sensitivity of depth regression losses to noise of larger objects. To bridge this gap, we comprehensively investigate regression and dice losses, examining their robustness under varying error levels and object sizes. 8/n

In this paper, we highlight this understudied problem of generalization to large objects. We find that modern frontal detectors struggle to generalize to large objects even on nearly balanced datasets like KITTI-360. 7/n

TLDR:.Monocular 3D detectors achieve great performance on cars and smaller objects. However, their performance drops on larger objects, leading to fatal accidents. Some attribute the failures to training data scarcity or the receptive field requirements of large objects. 6/n

✅ BEV Detectors work decent, but Dice loss in BEV segmentation boosts detection accuracy💡.✅ Dice loss is better than regression losses for large objects and large regression noise. 📊.✅ Result: SoTA on KITTI-360 Leaderboard and improves SoTA on nuScenes leaderboard. 🏁. 5/n

Key Findings: 💪.✅ Large object detection is a frontal/BEV representation and training loss problem.🚀.✅ All SoTA frontal detectors (including our GrooMeD-NMS, DEVIANT, and Cube R-CNN) get bad detection performance on large objects on a nearly balanced KITTI-360 dataset 💀. 4/n

Joint work with Yuliang Guo (Bosch Research North America), Xinyu Huang (Bosch Research North America), Liu Ren (Bosch Research North America) and Xiaoming Liu (MSU). 3/n.

Demo: MonoDETR is a frontal transformer head-based 3D detector that hardly detects large objects (buildings), while PBEV+SeaBird easily detects large objects. 2/n.

Check out the following #CVPR2024 work on improving the detection of large objects and making AVs safer. SeaBird: Segmentation in Bird's View with Dice Loss Improves Monocular 3D Detection of Large Objects. Code+Models: Paper: 1/n

RT @CVPR:

Grateful to be named as a top reviewer in #NeurIPS2023. Thanks to the organizers for this recognition and for complimentary registration.