Overview
Trajectory-Rig Decoupled Metric Geometry Learning (TRIG) is a geometry perception framework designed for autonomous driving. This framework addresses challenges in metric geometry and ego-motion estimation derived from synchronized multi-camera observations. TRIG operates by factorizing camera poses into two distinct components: the ego-trajectory and the camera-rig. This factorization facilitates the separate modeling of the vehicle's ego-motion and the static configuration of the multi-camera system.
Research Context
Vision-centric autonomous driving systems necessitate accurate estimations of metric geometry and ego-motion. These estimations are typically derived from synchronized observations captured by multiple cameras. While contemporary visual geometry models have demonstrated strong performance in tasks such as pose estimation, depth prediction, and 3D reconstruction, their design may not be optimally suited for multi-camera driving systems classified as rigid. A common limitation identified in these models is the encoding of camera poses as entangled representations. In such representations, time-varying ego-motion and static camera-rig geometry are modeled jointly. This joint modeling characteristic can limit the effective utilization of geometric priors inherent to the vehicle system.
Approach
TRIG's approach is centered on the factorization of camera poses into ego-trajectory and camera-rig components. This method allows for the independent modeling of ego-motion and the static multi-camera topology. Key elements of the methodology include:
- Decoupled Pose Encoding: This mechanism separately encodes the trajectory evolution and the rig geometry.
- Decoupled Supervision: Associated with the decoupled pose encoding, this supervision separately constrains trajectory evolution and rig geometry to promote metric-consistent learning.
- Sparse Temporal–Spatial Attention: TRIG incorporates a sparse Temporal–Spatial attention mechanism. This mechanism serves to differentiate between cross-camera interaction and temporal aggregation. By separating these aspects, the framework aims to reduce global attention costs while maintaining geometric reasoning capabilities.
Findings
Experiments conducted with TRIG on five distinct autonomous driving benchmarks yielded the following outcomes:
- TRIG achieved state-of-the-art performance in pose estimation.
- TRIG demonstrated state-of-the-art performance in metric depth prediction.
- TRIG attained state-of-the-art performance in 3D reconstruction.
Why This Matters
The ability of TRIG to achieve state-of-the-art performance in pose estimation, metric depth prediction, and 3D reconstruction for autonomous driving systems suggests an advancement in the precision and reliability of visual perception. By addressing the limitation of entangled representations in camera pose encoding, TRIG could enable more accurate and robust geometric understanding in complex driving environments.