Deep Reading Notes: CleanPose: Category-Level Object Pose Estimation via Causal Learning and Knowledge Distillation

1. Basic information

• Authors / affiliations: Xiao Lin et al.; Tongji University, State Key Laboratory of Autonomous Intelligent Unmanned Systems.
• Year / venue: 2025 / ICCV; reading mode: Deep reading.
• Paper: ICCV Open Access; code: GitHub.

2. Positioning in one sentence

CleanPose attributes RGB-D category-level pose generalization errors to hidden confounders and repeated / limited pose data. Keypoints act as mediators for front-door adjustment, followed by residual semantic distillation from a 3D foundation model.

2.1 Guide to the paper’s figures and tables

3. Problem setting and causal assumptions

• Inputs: An object RGB crop and point cloud back-projected using depth / intrinsics, typically with category labels / segmentation from a frontend.
• Outputs: Category-level 9DoF pose: Rotation, translation, and metric size.
• Causal graph: $\mathcal{X}\rightarrow\mathcal{M}\rightarrow\mathcal{Y}$ maps visual input through structural keypoints to pose. $\mathcal{U}$ affects both inputs and outputs, representing dataset/scene/pose-distribution bias.
• My analysis: Front-door adjustment assumes the mediator captures the relevant image-to-pose paths and that confounding effects on the mediator are sufficiently handled. These are modeling assumptions, not automatically guaranteed by the network.

4. Method details

Module	Input / output	Purpose
RGB-D encoder	DINOv2 image features + PointNet++ point features	Produce fused object features
Keypoint extractor	Object features → structural keypoint features	Construct causal mediators
Causal inference module	Current keypoints + memory-sampled features	Approximate front-door adjustment and suppress spurious correlations
Residual distillation	Student point features, ULIP-2 features	Add missing 3D semantic knowledge
Pose/size heads	Cleaned features	Regress 9DoF outputs

• Training requires the ULIP-2 teacher / sample queue. Testing remains a standard RGB-D pose pipeline without running the large teacher.

5. Experimental analysis

• REAL275, CAMERA25, and HouseCat6D use 3D IoU and rotation/translation success thresholds.
• Tables 1–5 compare main results, modules, and settings. The authors claim improvements over existing state of the art across multiple benchmarks.
• Evidence sufficiency: Multiple datasets support category-level claims. Further specificity tests for the causal module should reverse training-pose distributions, vary segmentation sources, and compare clean versus noisy depth.

6. Critical analysis and reproduction

• Strength: Separates added foundation-model information through residuals rather than treating it as a black-box feature boost. Causal and distillation modules address different biases.
• Limitations: Depends on RGB-D, masks, a teacher, and a memory queue. Real confounders are not explicitly annotated, making theoretical assumptions difficult to test fully.
• Reproduction priorities: ULIP-2 embedding alignment, front-door queue sampling, and DINOv2 / PointNet++ feature normalization.
• Improvements: Select memory samples using calibrated uncertainty, replace fixed-category teachers with multimodal language/shape priors, and test new categories rather than only new instances.

7. Research implications

• Feature strength is not the only category-level pose challenge. Data sampling, pose priors, and environmental co-occurrence can create shortcuts.

8. Additional original-paper results

Overall means support effectiveness, while category radar plots show uneven gains. Stable geometry and broad training coverage more readily benefit from debiased representations; variable or highly symmetric shapes still face multiple pose solutions. Qualitative examples can suggest reduced background shortcuts but cannot alone prove causal deconfounding. Module ablations and cross-environment improvements are the stronger tests.