Perez, Marc (Applus+ IDIADA, Institut de Robòtica i Informàtica Industrial CS), Agudo, Antonio (Institut de Robòtica i Informàtica Industrial CSIC-UPC), Dubbelman, Gijs (Eindhoven University of Technology), Jancura, Pavol (Eindhoven University of Technology)

Class Prototypical Loss for Enhanced Feature Separation in 3D Object Detection

Scheduled for presentation during the Poster Session "3D Object Detection" (FrAT13), Friday, September 27, 2024, 10:30−12:30, Foyer

2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada

This information is tentative and subject to change. Compiled on December 26, 2024

Abstract

We present a novel loss to increase the class separation of learned features for 3D object detection from lidar point clouds. To correctly classify objects, learned object-level feature distributions of each class need to be distinct. Therefore, we hypothesize that if we make the feature distributions of the classes more separated, then the overall performance of the object detector will improve. To this end, we calculate class prototypes as the mean and covariance of the feature vectors extracted from the annotated objects of each class. Then, we exploit these prototypes with a novel class prototypical loss, defined as the Mahalanobis distance from the feature vector of annotated objects to the corresponding class prototype. This auxiliary loss is then integrated with other object detection losses to improve the object-level feature separation between classes and the overall performance of the detector. We show results applying this loss to the NuScenes dataset where we get improvements of +3.85% and +1.76% mAP for 1 and 10 frames, respectively, compared to the baseline Centerpoint detector, while keeping the same inference computational cost.