$ show swc-17/SparseDrive
SparseDrive is an end-to-end autonomous driving framework that processes multi-view camera images to perform unified perception and planning. It simultaneously predicts 3D object bounding boxes with tracking IDs, vectorized HD map elements (dividers, boundaries, pedestrian crossings), agent motion trajectories, and ego vehicle planning trajectories. The architecture employs a ResNet-50 FPN backbone, deformable feature aggregation for multi-view fusion, instance banks for temporal object representation, and specialized heads for each task. Key innovations include hierarchical planning with collision-aware rescore and symmetric sparse representation for detection, mapping, and motion tasks. The system is designed for efficiency, achieving 9.0 FPS on NVIDIA A100 with significantly reduced training time compared to prior art like UniAD.
The LLM identified 2 files worth highlighting for reuse:
Implements a CUDA-accelerated deformable feature aggregation module for fusing multi-view camera features. Uses deformable attention to sample features from multiple views based on learned offsets, enabling efficient and effective feature fusion for bird's-eye view perception. Self-contained with clear Python/CUDA interface, reusable in other multi-view architectures.
Maintains a bank of instance features and metadata (bounding boxes, confidence) across time steps. Updates bank with new detections and applies confidence decay for object disappearance. Enables consistent object tracking and temporal reasoning in video-based perception. Modular design allows reuse in other tracking or video object detection frameworks with minimal adaptation.