Research map / Point clouds / Segmentation and detection
itKD: Interchange Transfer-Based Knowledge Distillation for 3D Object Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add interchange transfer loss L_it (interchange reconstruction via shared autoencoder) (ablation) | channel-wise autoencoder KD on student (1/4) · Waymo validation | mAPH L2 average: 50.72 -> 52.13 (+1.41) | Helped |
| add compressed representation loss L_cr (ablation) | channel-wise autoencoder KD on student (1/4) · Waymo validation | mAPH L2 average: 52.13 -> 53.07 (+0.94) | Helped |
| add head relation-aware self-attention loss L_attn (ablation) | student (1/4) with autoencoder KD · Waymo validation | mAPH L2 average: 53.07 -> 53.54 (+0.47) | Helped |
| interchange reconstruction vs self-reconstruction in the autoencoder (ablation) | autoencoder KD without head attention loss · Waymo validation | mAPH L2 average: 51.37 -> 53.07 | Helped |
| share autoencoder parameters between teacher and student (ablation) | channel-wise autoencoder · Waymo validation | mAPH L2 average: 50.11 -> 53.07 | Helped |
| upsample student feature to teacher size (S→T) as buffer vs downsample teacher (T→S) or median (S+T)/2 (ablation) | buffer layer for channel mismatch · Waymo validation | mAPH L2 average: 53.07 vs 53.05 vs 52.95 | No clear effect |
| direct KL loss and l1 loss applied to all detection heads (ablation) | student (1/4) multi-head KD · Waymo validation | mAPH L2 average: KL 49.50 and l1 49.82 vs student 50.72 | Hurt |
| head relation-aware self-attention loss L_attn vs AT and inter/intra-only losses (ablation) | multi-head KD without autoencoder · Waymo validation | mAPH L2 average: AT 51.52, L_inter 51.40, L_intra 51.87, L_attn 52.08 | Helped |
| itKD distillation on student (1/4) vs no KD, Waymo | CenterPoint student (1/4) · Waymo validation | mAPH L2 average: 50.72 -> 53.54 | Helped |
| itKD on student (1/4) vs strongest baselines on Waymo vehicle L2 | CenterPoint student (1/4) · Waymo validation | mAPH L2 vehicle: 58.81 vs Obj. DGCNN 58.55 | Helped |
| itKD on student (1/4) vs strongest baselines on nuScenes | CenterPoint student (1/4) · nuScenes validation | NDS: 53.90 vs SparseKD 53.01 | Helped |
| itKD on student (1/2) width | CenterPoint student (1/2) · Waymo validation | mAPH L2 average: 59.04 vs teacher 60.13 with 3.5x fewer params | Helped |
Closest papers by meaning
- PointDistiller: Structured Knowledge Distillation Towards Efficient and Compact 3D Detection (CVPR 2023 · similarity 0.89)
- Towards Efficient 3D Object Detection with Knowledge Distillation (NeurIPS 2022 · similarity 0.89)
- DCSF-KD: Dynamic Channel-wise Spatial Feature Knowledge Distillation for Object Detection (AAAI 2025 · similarity 0.86)
- CrossKD: Cross-Head Knowledge Distillation for Object Detection (CVPR 2024 · similarity 0.86)
- 3D Feature Prediction for Masked-AutoEncoder-Based Point Cloud Pretraining (ICLR 2024 · similarity 0.86)
- One for All: Multi-Domain Joint Training for Point Cloud Based 3D Object Detection (NeurIPS 2024 · similarity 0.86)
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.