Transfer Learning from Simulated to Real Scenes for Monocular 3D Object Detection
Authors/Creators
-
1.
University of Cagliari
-
2.
Technical University of Munich
-
3.
University of California, Merced
- 4. SETLabs Research GmbH, Munich, Germany
-
5.
Universidad de Las Palmas de Gran Canaria
-
6.
University of California, San Diego
Description
Accurately detecting 3D objects from monocular images in dynamic roadside scenarios remains a challenging problem due to varying camera perspectives and unpredictable scene conditions. This paper introduces a two-stage training strategy to address these challenges. Our approach initially trains a model on the large-scale synthetic dataset, RoadSense3D, which offers a diverse range of scenarios for robust feature learning. Subsequently, we fine-tune the model on a combination of real-world datasets to enhance its adaptability to practical conditions. Experimental results of the Cube R-CNN model on challenging public benchmarks show a remarkable improvement in detection performance, with a mean average precision rising from 0.26 to 12.76 on the TUM Traffic A9 Highway dataset and from 2.09 to 6.60 on the DAIR-V2X-I dataset, when performing transfer learning. Code, data, and qualitative video results are available at https://roadsense3d.github.io.
Files
Abstract.pdf
Files
(120.6 kB)
| Name | Size | Download all |
|---|---|---|
|
md5:ba9af8508363c73e74f01d709b813871
|
120.6 kB | Preview Download |