Source-linked AI summary
Density-aware Chamfer Distance as a Comprehensive Metric for Point Cloud Completion
Tong Wu, Liang Pan, Junzhe Zhang, Tai Wang, Ziwei Liu, Dahua Lin
TL;DR
Existing point-cloud metrics can overlook local density, detailed structures, or outlier effects, making consistent evaluation difficult. The paper introduces DCD, a bounded CD-derived measure incorporating density distribution, and evaluates it for completion, training, and guided down-sampling. DCD provides more reliable evaluations, improves results across three metrics when used as a loss, and supports superior visual quality, while broader task generalization remains to be tested.
Problem
CD is insensitive to mismatched density and outliers, while EMD emphasizes global distribution and overlooks detailed structures; both can yield inconsistent evaluations.
Method
DCD modifies CD with query-frequency and Taylor-expansion terms, and the paper adds a point discriminator guiding priority-based down-sampling.
Results
DCD provides more reliable evaluation when CD and EMD contradict, while DCD-trained models reduce DCD, EMD, and CD and the balanced design achieves superior visual quality.
Takeaways & Limitations
DCD offers a bounded, computationally efficient similarity measure that considers density distribution and detailed structures for point-cloud completion evaluation and training.
Takeaways & Limitations
The paper investigates DCD primarily on point-cloud completion and leaves validation of generalization across other tasks for future work.
Abstract
from arXiv · showhide
Chamfer Distance (CD) and Earth Mover's Distance (EMD) are two broadly adopted metrics for measuring the similarity between two point sets. However, CD is usually insensitive to mismatched local density, and EMD is usually dominated by global distribution while overlooks the fidelity of detailed structures. Besides, their unbounded value range induces a heavy influence from the outliers. These defects prevent them from providing a consistent evaluation. To tackle these problems, we propose a new similarity measure named Density-aware Chamfer Distance (DCD). It is derived from CD and benefits from several desirable properties: 1) it can detect disparity of density distributions and is thus a more intensive measure of similarity compared to CD; 2) it is stricter with detailed structures and significantly more computationally efficient than EMD; 3) the bounded value range encourages a more stable and reasonable evaluation over the whole test set. We adopt DCD to evaluate the point cloud completion task, where experimental results show that DCD pays attention to both the overall structure and local geometric details and provides a more reliable evaluation even when CD and EMD contradict each other. We can also use DCD as the training loss, which outperforms the same model trained with CD loss on all three metrics. In addition, we propose a novel point discriminator module that estimates the priority for another guided down-sampling step, and it achieves noticeable improvements under DCD together with competitive results for both CD and EMD. We hope our work could pave the way for a more comprehensive and practical point cloud similarity evaluation. Our code will be available at: https://github.com/wutong16/Density_aware_Chamfer_Distance .
1 Introduction
Point-cloud similarity metrics must handle unordered, irregular data with varying point counts while supporting fair evaluation and training. The paper proposes DCD to address CD’s density blindness and outlier sensitivity, reporting more reliable evaluation and improved completion results.
- Point-cloud similarity measures are important for training and fair evaluation but must accommodate unordered, irregular data with varying point numbers.
- CD is efficient and flexible but can miss mismatched density distributions and become strongly affected by outliers.
- DCD extends CD with query-frequency and Taylor-expansion terms to detect density disparities and tolerate outliers.
- DCD follows EMD-like trends across varying point distributions while remaining more computationally efficient and capturing details more effectively.
- DCD provides more consistent evaluation when CD and EMD disagree and also handles non-uniform ground-truth sampling such as curvature-based sampling.
- Using DCD as a loss lowers DCD, EMD, and CD versus CD-trained models, while guided down-sampling improves DCD with competitive CD and EMD results.
2 Related Works
Related work covers point-cloud completion methods and established point-set distances. It positions DCD as an explicit distance that retains EMD-like behavior while better detecting detail-preservation issues.
- Point Cloud Completion: Point-cloud completion methods recover complete shapes from partial observations, progressing from voxel representations to raw-point encoder-decoder and attention-based approaches.
- Point Cloud Distance: CD measures nearest-neighbour distances efficiently for point sets with different sizes, whereas EMD finds a one-to-one mapping but is much more computationally expensive.
- Point Cloud Distance: DCD shares EMD’s properties in many cases while detecting detail-preservation issues more effectively.
3 Density-aware Chamfer Distance for Point Sets
DCD extends nearest-neighbour Chamfer Distance by incorporating point-specific query frequency and bounded distance transformations. It is designed to detect density mismatch, reduce outlier influence, preserve local detail, and offer a more efficient alternative to EMD.
- Preliminaries: CD averages bidirectional nearest-neighbour distances, whereas EMD computes distances after an expensive one-to-one mapping between equally sized point sets.EMD’s assigned pairs may be far apart, weakening their physical meaning.
- Motivation: CD can overlook mismatched density because points querying the same neighbour remain unaware of one another.The contribution of a shared neighbour is not reduced as its query frequency increases.
- Bounded formulation: Taylor approximation maps point-level distances into [0, 1], bounding the overall shape distance and reducing the influence of large outliers.The temperature scalar α adjusts sensitivity to the distance distribution.
- Comparisons: Figure 2 contrasts DCD’s locally assigned nearest neighbours with EMD’s global one-to-one assignment, while Figure 3 compares loss and gradient curves across CD variants and DCD settings.The DCD curves are examined under different query frequencies n and temperature values α.
- Density-aware formulation: DCD normalizes each point’s contribution with the reciprocal query frequency 1/n_y, making the metric sensitive to local density distributions.It is derived from CD and uses point-specific query frequency to measure distribution quality.
4 Incorporating Balanced Design in Point Cloud Completion
The framework combines a coarse-to-fine completion pipeline with density-aware training and a point discriminator that guides down-sampling to preserve balanced, detailed outputs.
- Point Cloud Completion Framework: The two-stage pipeline generates a coarse complete point cloud, then adds local geometric features to produce a more detailed final output.The first stage uses global features, while the second incorporates local features and preserves fidelity to the input.
- Balanced Design: CD can favor placing more points in confidently observed regions even when this creates density imbalance and conflicts with human assessment.The paper identifies this imbalance as a central motivation for its balanced design.
- Balanced Design: The method replaces LCD with a density-aware loss and uses a point discriminator to estimate each point’s importance for guided down-sampling.The discriminator output supplies sampling priority in the final down-sampling stage.
- Point Discriminator: The target function g(x) combines query frequency and distance to the ground-truth surface, mapping point cases into a suitable learning range.For n > 0 it uses a logarithmic frequency term; for n = 0, scaled distance helps distinguish over-populated surface points from distant points.
- Point Discriminator: An MLP predicts g(x) from local, global, and coordinate features, producing a scalar used to determine sampling privilege during inference.The module aggregates neighboring-point information through edge convolution before regression.
- Guided Down-sampling: During inference, discriminator-guided down-sampling is combined with FPS to reduce unreasonable points while maintaining a predefined output size.The process is designed to mitigate FPS’s tendency to select outer-region points and include outliers.
5 Experiments
Experiments on MVP compare completion methods and metrics, showing that DCD better captures balanced distributions and local structures while providing more stable evaluation. DCD-based training and guided down-sampling further improve results across the reported metrics.
- 5.1 Comparison of the Methods: On MVP, the proposed method achieves the lowest DCD, second-lowest CD, and comparable EMD with VRC-EMD.Qualitative results show superior global point-distribution balance and local structures.
- 5.2 Comparison of the Metrics: DCD decreases only when both CD and EMD are relatively low, addressing their inconsistent rankings when the metrics disagree.The test-set scatter plot shows weak positive CD–EMD correlation and lower DCD toward jointly low CD and EMD.
- 5.2 Comparison of the Metrics: DCD’s [0,1] range supports more comprehensive and stable test-set statistics than unbounded CD and EMD.For CD, the top 50% of shape-level losses contribute 80% of normalized accumulation, while the top 25% contribute 50%.
- 5.3 Ablation Study: The ablation study combines DCD loss, point up-sampling, and guided down-sampling to obtain the lowest EMD and DCD with relatively low CD.Point up-sampling lowers EMD but raises CD, while guided down-sampling improves the combined design.
- 5.4 Performance as the Loss Function: Training with LDCD produces the lowest DCD, reduces EMD, and can slightly reduce CD compared with LCD-trained models.Its time consumption is comparable with LCD, unlike the heavier computational burden associated with LEMD training.
6 Conclusion
The paper concludes that DCD is a bounded, computationally efficient point-cloud similarity measure that considers density distribution and detailed structures. Experiments report improved DCD performance and superior visual quality over previous methods.
- DCD is bounded, computationally efficient, and designed to reflect both density distribution and detailed structures.
- The proposed method achieves noticeable improvements under DCD and superior visual quality compared with previous works.
A Further Analysis of Balanced Chamfer Distance
Further analysis examines the temperature scalar α used in DCD and extensions for point sets with unequal sizes. The reported settings balance numerical behavior, training practicality, and one-to-many point matching.
- A.1 The consistency and choice of hyper-parameter α: Setting α = 1000 places typical nearest-pair distances near the exponential term’s high-gradient region and yields the largest observed variance.
- A.1 The consistency and choice of hyper-parameter α: DCD’s relative values and trends remain consistent across α under different data distributions, although absolute values change.The same α should be used across methods for fair evaluation.
- A.2 DCD with Different Point Numbers: For unequal point-set sizes, the naive DCD extension uses η or 1/η to represent one-to-many mapping.
- A.2 DCD with Different Point Numbers: A second unequal-size variant replaces 1/n̂_y with max(η/n̂_y, 1) to avoid negative first-term contributions and preserve queried-point contribution.
- A.2 DCD with Different Point Numbers: Equation 8 is used during training between a 1024-point coarse shape and a 2048-point ground truth.
A.3 Time Complexity.
The paper contrasts EMD’s assignment-based computational cost with the nearest-neighbor computation used by CD and DCD. Experiments report that CD and DCD are significantly faster than EMD, whose runtime also varies with point distributions.
- A.3 Time Complexity.: EMD uses iterative linear assignment with practical time complexity between O(n^2) and O(n^3), typically requiring O(n^2) memory.
- A.3 Time Complexity.: The supplementary analysis visualizes how α affects the exponential function, DCD statistics, PCN training loss, and per-instance evaluation results.
- A.3 Time Complexity.: CD and DCD rely mainly on nearest-neighbor selection, usually O(n) in time and accelerable with structures such as KD-trees.
- A.3 Time Complexity.: CD and DCD are significantly more efficient to compute than EMD, while EMD runtime increases with the difficulty of finding an optimal assignment.
A.4 Evaluation on other tasks and ground truth distributions.
DCD extends evaluation beyond uniform-density assumptions by measuring fidelity to ground-truth distributions, including non-uniform sampling, while avoiding strong geometric assumptions.
- Evaluation on other tasks and ground truth distributions: DCD is also presented as suitable for point cloud upsampling and denoising when ground truth with desirable point distributions is available.These tasks require outputs that are uniform, clean, and faithfully located on the underlying surface.
- Evaluation on other tasks and ground truth distributions: DCD focuses on faithfulness to the ground-truth distribution without requiring strong assumptions about local surfaces or expected point counts.Unlike existing metrics that encourage uniformity, DCD can reflect specific non-uniform distributions such as curvature-based sampling.
- Evaluation on other tasks and ground truth distributions: As curvature-based sampling increases, DCD and EMD reflect increasingly mismatched density, whereas CD and uniform loss remain insensitive.
- Evaluation on other tasks and ground truth distributions: DCD’s per-point distance visualization reflects local density mismatches between point sets more effectively than CD.
- Effect of hyper-parameters in LDCD: Larger α usually lowers DCD, while λ controls a trade-off among CD, EMD, and DCD performance.λ = 0 usually gives the best CD but sub-optimal EMD and DCD; λ = 1 hurts CD.
B User Study on Visual Quality
A user study compares completion outputs and metric-selected results against human judgments of global shape, local detail, and point distribution. The proposed method receives the strongest preference, and DCD aligns most closely with human vision.
- User study design: 25 partial inputs and five methods produced 125 completion outputs for a user study evaluating comprehensive visual quality.The study assessed global-shape similarity, local-detail fidelity, and point-cloud distribution quality.
- User study design: The compared methods were chosen because each achieved favorable results under at least one metric, creating a basis for evaluating metric quality.
- User study design: 15 volunteers selected up to two outputs per shape as having the highest comprehensive score.
- Comparison of the methods: The proposed method achieved the highest average favorite-selection ratio and outperformed the other methods by a large margin.The authors interpret this as evidence of high visual quality.
- Comparison of the metrics: DCD achieved the highest alignment with human vision when selecting the top-1 completion result for each shape.Alignment was measured by the average volunteer-selection ratio of the result chosen by each metric.
C.3 Training Details
The two-stage framework combines point-cloud distance losses with VAE regularization and point-discriminator training, using specified loss weights and optional DCD-training adjustments.
- Training losses: The full training loss combines discriminator regression loss Lh, paired point-cloud distance loss Ld, and KL-divergence loss LKL.Ld contains multiple paired point-cloud distances, while LKL regularizes the dual-path VAE architecture.
- DCD-based training: The framework retains Chamfer Distance as its training loss in the described implementation, with LDCD replacing occurrences of LCD when adopted.
- Training losses: The KL term regularizes latent distributions through divergences involving encoders qφ and qψ, weighted by λKL.
- Training configuration: The experiments set λ1 = 10, λ2 = 0.5, λ3 = 1, and λKL = 20.
- DCD-based training: When using LDCD, different α values can be assigned across loss terms, and an additional L1-version CD term can be added.The reported implementation uses α = 50 for the first term and α = 100 for the others, together with CD-P.
D Limitations and Future Work
The paper primarily investigates DCD on point cloud completion, leaving its generalization across other tasks for future validation.
- Limitations and Future Work: The main limitation is that DCD’s properties are investigated on point cloud completion rather than across multiple tasks.The authors identify validation of generalization ability across tasks as future work.
- Limitations and Future Work: The paper presents DCD as applicable to other tasks and scenarios, but does not yet provide the planned cross-task generalization experiments.