Research map / Detection and segmentation / Weakly supervised detection and segmentation
HINTED: Hard Instance Enhanced Detector with Mixed-Density Feature Fusion for Sparsely-Supervised 3D Object Detection
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 HINTED, a teacher-student framework for sparsely-supervised 3D object detection that uses a self-boosting teacher to generate more pseudo-labels and a mixed-density student that fuses regular and locally down-sampled BEV features to focus on hard instances.
Problem
Sparsely-supervised 3D object detection relies on high-threshold pseudo-labels, which miss hard instances (distant, incomplete point clouds) and leave them with inadequate positive supervision.
Approach
A self-boosting teacher updates pseudo-labels on sparsely labeled scenes by NMS-style filtering against the ground-truth instance, and a mixed-density student fuses features from a regular view and a nearby-downsampled view with adaptive SE-style weights, enforcing consistent predictions across densities to boost hard-instance learning.
Key results
- On KITTI val with 2% sparse labels, HINTED reaches 94.3/82.5/78.7 car AP and 66.5/59.9/53.7 pedestrian AP (easy/mod/hard), versus CoIn++ at 92.0/79.5/71.5 car and 46.7/36.1/31.2 pedestrian.
- HINTED's cyclist AP (94.6/76.3/73.0) exceeds fully-supervised VoxelRCNN's cyclist AP (88.7/72.5/68.2).
- In the semi-supervised setting with 2% labels on PV-RCNN, HINTED achieves 60.5 avg AP at 1% and 70.8 at 2%, versus HSSDA's 59.5 and 68.6.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the self-boosting teacher to the teacher-student framework (ablation) | HINTED on KITTI val split, Car Easy · KITTI val split, Car Easy | 3D AP: 89.5 -> 93.8 | Helped |
| Add mixed-density features B and B− to the SBT model (ablation) | HINTED on KITTI val split, Car Easy · KITTI val split, Car Easy | 3D AP: 93.8 -> 94.3 | Helped |
| Use local random sampling at 20% rather than other tested sampling strategies and rates (ablation) | HINTED · KITTI val split, Car Mod. | 3D AP: 81.67; 81.71; 81.59; 81.66; 82.56; 82.34 | Mixed |
| Replace attention-based fusion with ROI-based feature fusion (ablation) | HINTED · KITTI val split, Car Mod. | 3D AP: 82.3 -> 77.4 | Hurt |
| Replace attention-based fusion with average-based feature fusion (ablation) | HINTED · KITTI val split, Car Mod. | 3D AP: 82.3 -> 80.3 | Hurt |
| Set γ1, γ2, γ3 to 0.1, 1, 1 (ablation) | HINTED · KITTI val split, mean of three classes at Mod. | 3D AP: 72.9 | Helped |
| Use HINTED instead of CoIn++ | VoxelRCNN sparsely-supervised detection · KITTI val split, Pedestrian category | 3D AP | Helped |
| Use HINTED instead of CoIn++ | VoxelRCNN sparsely-supervised detection · KITTI val split, Cyclist category | 3D AP | Helped |
| Use HINTED instead of HSSDA | PV-RCNN semi-supervised detection with 2% labels · KITTI val split, Avg. | 3D AP: 68.6 -> 70.8 | Helped |
| Use HINTED instead of HSSDA | PV-RCNN semi-supervised detection with 1% labels · KITTI val split, Avg. | 3D AP: 59.5 -> 60.5 | Helped |
Surprising findings
- Sparse-supervised HINTED outperforms fully-supervised VoxelRCNN on the cyclist category.
Limitations
- Training is slower because feature maps for the student's two views cannot be computed in parallel with shared weights due to memory limits.
- Gains on car category are absent in the semi-supervised setting, possibly because focusing on hard instances loses simple structural features.
- Evaluated only on KITTI.
Experimental setup
Datasets: KITTI
Benchmarks: KITTI
Metrics: 3D AP R40
Baselines: VoxelRCNN SS3D CoIn CoIn++ 3DIoUMatch DetMatch HSSDA WS3D WSS3D
Builds on: CoIn Mean Teacher SE block VoxelRCNN PV-RCNN
Closest papers by meaning
- SP3D: Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts (CVPR 2025 · similarity 0.87)
- UNION: Unsupervised 3D Object Detection using Object Appearance-based Pseudo-Classes (NeurIPS 2024 · similarity 0.86)
- As Pseudo-Label Free as Possible: Leveraging Adaptive Feature Generation for Sparsely Annotated Object Detection (AAAI 2025 · similarity 0.85)
- Semi-supervised 3D Object Detection with PatchTeacher and PillarMix (AAAI 2024 · similarity 0.85)
- Leveraging Imagery Data with Spatial Point Prior for Weakly Semi-supervised 3D Object Detection (AAAI 2024 · similarity 0.85)
- Pseudo Label Refinery for Unsupervised Domain Adaptation on Cross-dataset 3D Object Detection (CVPR 2024 · similarity 0.85)
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.