Research map / Point clouds / Segmentation and detection
iDet3D: Towards Efficient Interactive Object Detection for LiDAR Point Clouds
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes iDet3D, a point-based interactive 3D object detector that takes sparse 2D clicks on LiDAR scenes and detects multiple multi-class objects, using negative click simulation, dense click guidance, and spatial click propagation.
Problem
Annotating many multi-class 3D objects in LiDAR point clouds is costly; existing semi-automatic annotation tools handle only one object per interaction and struggle with false positives in sparse point clouds.
Approach
Extends a single-stage point-based detector (IA-SSD) with 2D click encodings, negative clicks sampled from background points with high foreground scores, dense click fusion at every downsampling layer, and cosine-similarity click prototypes that propagate clicks to other same-class objects.
Key results
- On KITTI val with five clicks, iDet3D (IA-SSD backbone) reaches 98.55 / 90.37 / 90.27 AP for Car (easy/mod/hard) and 70.07 / 65.74 / 60.57 for Pedestrian.
- On nuScenes val with five clicks, iDet3D achieves 51.83 mAP versus 45.0 for SASA and 42.66 for 3DSSD.
- In a user study, iDet3D averaged 2.5 clicks per scene versus 19 for LATTE and 45 for manual annotation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DCG to the vanilla model (ablation) | iDet3D with IA-SSD backbone · KITTI val, Car Moderate (IoU=0.7) | AP: 82.08 -> 88.86 | Helped |
| Add DCG to the vanilla model (ablation) | iDet3D with IA-SSD backbone · KITTI val, Pedestrian Easy (IoU=0.5) | AP: 58.76 -> 60.39 | Helped |
| Add NCS to the model with DCG (ablation) | iDet3D with IA-SSD backbone · KITTI val, Pedestrian Easy (IoU=0.5) | AP: 60.39 -> 69.94 | Helped |
| Add NCS to the model with DCG (ablation) | iDet3D with IA-SSD backbone · KITTI val, Cyclist Easy (IoU=0.5) | AP: 89.00 -> 98.46 | Helped |
| Add SCP to the model with DCG and NCS (ablation) | iDet3D with IA-SSD backbone · KITTI val, Cyclist Moderate (IoU=0.5) | AP: 84.60 -> 88.00 | Helped |
| Add SCP to the model with DCG and NCS (ablation) | iDet3D with IA-SSD backbone · KITTI val, Cyclist Easy (IoU=0.5) | AP: 98.46 -> 98.43 | Hurt |
| Add all three components: DCG, NCS, and SCP (ablation) | Vanilla iDet3D with IA-SSD backbone · KITTI val, Car Easy (IoU=0.7) | AP: 88.76 -> 98.55 | Helped |
| Increase user clicks from 1 to 5 | iDet3D with IA-SSD backbone · nuScenes validation | mAP: 41.64 -> 51.83 | Helped |
| Use iDet3D instead of the vanilla interactive model | IA-SSD backbone · nuScenes validation | mAP: 39.48 -> 51.83 | Helped |
Surprising findings
- Vanilla interactive model with positive clicks alone gives little gain over the non-interactive backbone, indicating that false positives and diluted click guidance limit it.
Limitations
- Analyzes only single LiDAR frames; multi-frame scenarios are not handled.
- Simulated clicks in main experiments; user study is small and reported with coarse time bounds.
- nuScenes numbers from re-implemented IA-SSD are not directly comparable with reported baselines.
Experimental setup
Datasets: KITTI nuScenes
Benchmarks: KITTI val nuScenes val
Metrics: AP (3D, IoU 0.7 Car / 0.5 Ped and Cyc) mAP precision recall clicks per instance clicks per scene time per scene
Baselines: VoxelNet SECOND PointPillars Part-A2 3DSSD IA-SSD SASA LATTE Vanilla iDet3D
Builds on: IA-SSD 3DSSD interactive segmentation click simulation object counting prototype matching
Closest papers by meaning
- GBlobs: Explicit Local Structure via Gaussian Blobs for Improved Cross-Domain LiDAR-based 3D Object Detection (CVPR 2025 · similarity 0.88)
- LiDAR-Net: A Real-scanned 3D Point Cloud Dataset for Indoor Scenes (CVPR 2024 · similarity 0.88)
- MixSup: Mixed-grained Supervision for Label-efficient LiDAR-based 3D Object Detection (ICLR 2024 · similarity 0.88)
- ImOV3D: Learning Open Vocabulary Point Clouds 3D Object Detection from Only 2D Images (NeurIPS 2024 · similarity 0.87)
- STONE: A Submodular Optimization Framework for Active 3D Object Detection (NeurIPS 2024 · similarity 0.87)
- PTT: Point-Trajectory Transformer for Efficient Temporal 3D Object Detection (CVPR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.