Research map / Continual learning / Class-incremental
Incremental Object Keypoint Learning
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.
Introduces Incremental object Keypoint Learning (IKL), where new keypoints are learned from data labeled only with new keypoints without retaining old data, and proposes the exemplar-free two-stage KAMP baseline that uses a KA-Net and a keypoint-oriented spatial distillation loss.
Problem
Pretrained keypoint estimators cannot detect newly defined keypoints at test time, and retraining with all keypoint labels is costly; incremental learning for keypoint regression with only new-keypoint labels (label non-co-occurrence) has not been studied.
Approach
A two-stage scheme: Stage-I trains an auxiliary KA-Net to predict an old keypoint from nearby new keypoint heatmaps and visual features (pseudo-labeled by the old model); Stage-II distills old keypoints with a spatial-softmax keypoint-oriented distillation loss plus KA-Net distillation while learning new keypoints, enabling positive transfer to old keypoints.
Key results
- On Split MPII, KAMP reaches AAA4 of 79.93 versus 75.75 for LWF, with positive average transfer AT4 of 1.80.
- On Split Head-2023 under 1-shot, KAMP achieves MRE 4.35 versus 5.01 for EGT.
- KAMP is the only method among compared exemplar-free methods to consistently achieve the highest AT and MT across all four datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply KAMP instead of LWF | HRNet · Split MPII | AAA4: 75.75 -> 79.93 | Helped |
| apply KAMP instead of the strongest listed baseline | HRNet · Split Head-2023 | A-MRE4: 3.71 -> 2.32 | Helped |
| apply KAMP instead of the strongest listed baseline | HRNet · Split ATRW | AAA3: 89.34 -> 93.16 | Helped |
| apply KAMP instead of the strongest listed baseline | HRNet · Split Chest | A-MRE1: 6.35 -> 5.67 | Helped |
| replace LWF's keypoint-wise normalization with keypoint-oriented spatial distillation ℓKSD (ablation) | KAMP on Split MPII · Split MPII | AAA4: 75.75 -> 76.93 | Helped |
| add a randomly constructed KA-Net task (ablation) | KAMP on Split MPII · Split MPII | AAA4: 76.93 -> 77.13 | Helped |
| use physical knowledge to construct the KA-Net task instead of random construction (ablation) | KAMP on Split MPII · Split MPII | AAA4: 77.13 -> 79.93 | Helped |
| incrementally apply KAMP to MetaPoint+ | MetaPoint+ on Split MPII · Split MPII | PCK [10]: 65.71 -> 73.49 | Helped |
Surprising findings
- Traditional CSS/CIL distillation methods (SFD, WF, GBD) get strong transfer metrics but poor overall accuracy because they become too rigid to learn new keypoints.
- Old keypoints can improve beyond their pre-IKL performance (positive transfer) rather than only being preserved.
Limitations
- Only tested on keypoints within the same object category; extension to novel categories is future work.
- Experiments use small datasets (under 400 images for medical splits) and simulated incremental splits.
- KA-Net relies on proximity-based auxiliary task creation and a single chosen tuple per step.
Experimental setup
Datasets: Head-2023 (Cephalometric, MICCAI 2023) Chest MPII ATRW
Benchmarks: Split Chest Split Head-2023 Split MPII Split ATRW
Metrics: MRE PCK average accuracy (AAA) average transfer (AT) maximal transfer (MT)
Baselines: EWC RW MAS LWF AFEC CPR SFD WF GBD CC2D EGT UKL
Builds on: LWF HRNet CAPE CHF
Closest papers by meaning
- KeyPose: Category-Level 6D Object Pose Estimation with Self-Adaptive Keypoints (AAAI 2025 · similarity 0.85)
- ESCAPE: Encoding Super-keypoints for Category-Agnostic Pose Estimation (CVPR 2024 · similarity 0.85)
- KptLLM: Unveiling the Power of Large Language Model for Keypoint Comprehension (NeurIPS 2024 · similarity 0.85)
- Meta-Point Learning and Refining for Category-Agnostic Pose Estimation (CVPR 2024 · similarity 0.85)
- Unsupervised 3D Structure Inference from Category-Specific Image Collections (CVPR 2024 · similarity 0.84)
- Instance-Adaptive and Geometric-Aware Keypoint Learning for Category-Level 6D Object Pose Estimation (CVPR 2024 · similarity 0.83)
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.