Occupancy-Assisted Surround-View Images 3D Object Detection
In the modern field of surround-view 3D detection, there has been a significant increase in interest. Existing methods are mainly focused on constructing dense Bird’s Eye View(BEV) features or utilizing sparse queries for detection. Additionally, there is a popular occupancy task aimed at...
Saved in:
| Main Author: | |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10599245/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | In the modern field of surround-view 3D detection, there has been a significant increase in interest. Existing methods are mainly focused on constructing dense Bird’s Eye View(BEV) features or utilizing sparse queries for detection. Additionally, there is a popular occupancy task aimed at predicting the semantic occupancy of 3D voxel space. In this paper, I innovatively improve the common 3D detection pipeline by leveraging the 3D positional and semantic information extracted from the semantic occupancy task. Specifically, I first generate a voxel feature and perform occupancy prediction. I use the results of occupancy prediction, which include the coordinates, categories, and features of the occupied voxels, to generate a series of occupied queries enriched with semantic and positional information. Then, I compress the voxel feature in height to obtain BEV features and use the coordinates of the occupied voxels to determine the occupied BEV queries, updating them through the aggregation of image features. Additionally, when performing temporal modeling, I only interact with the occupied BEV features. With the help of the occupancy task, I efficiently improve the existing 3D detection pipeline and achieve state-of-the-art results on the nuScenes test set, with an mAP of 61.4% and NDS of 52.6%, gaining 1.8% mAP and 1.5% NDS compared to MV2D. Code will be available. |
|---|---|
| ISSN: | 2169-3536 |