PaperlayerView exact revision on arXiv

Source-linked AI summary

BRICE: Brain-inspired Recurrent Illusory Contour Encoder for Point Cloud Completion

Yunlei Wang, Dawei Li, Xue-song Tang

arXiv:2610.23001v1cs.CV

TL;DR

Point-cloud completion methods often rely on deep attention stacks and feedforward-only propagation. BRICE uses predictive-coding recurrent feedback, and experiments report improvements across mainstream baselines and datasets, though two of seven PCN baselines decline.

  • Problem

    Existing point-cloud completion methods often stack deep attention modules and rely on feedforward-only propagation, while lacking rigorous theoretical guidance.

    Source: Introduction · p. 2

  • Method

    BRICE is a predictive-coding recurrent encoder that combines feedforward processing with feedback from deeper layers for point-cloud completion.

    Source: Introduction · p. 2, Predictive coding for updating · p. 4

  • Results

    BRICE generally improves completion across multiple datasets and baselines, though on PCN it improves L1-CD for five of seven baselines and declines for two.

    Source: Results · pp. 6–7

  • Takeaways & Limitations

    BRICE is currently suited to completion baselines with clear encoder-decoder architectures; its applicability to other architectures remains unexplored.

    Source: Conclusion · p. 7

    BRICE's error-correction account assumes natural-signal variation affects point clouds through hidden physical causes.

    Source: Predictive coding for updating · p. 4

Abstract

from arXiv · show

Mainstream point cloud completion methods are unanimously based on deep learning architectures. On one hand, these methods tend to repetitively stacking attention modules to enhance network performance, which drifts away from the neat design of unified and interpret-able networks; on the other hand, nearly all existing architectures rely solely on feed-forward connections to propagate information, neglecting feedback from deeper layers. The recurrent neural architecture, formed by feedforward-feedback loops, have shown its ability to generate Illusory Contours (ICs) in the V1 brain region of mammals; and ICs are essentially a completion mechanism of the brain for visual information. Motivated by this, we design a Brain-inspired Recurrent Illusory Contour Encoder (BRICE) to "imagine" what is missing in the incomplete 3D point clouds. Building upon the foundation model of Predictive Coding (PC), the plug-and-play nature enables BRICE to be seamlessly integrated into mainstream point cloud completion baselines. Experimental results show that BRICE can enhance various completion baselines on different 3D datasets such as MVP2K, PCN, and PlantPCom. Our code: Https://github.com/wangyunlei-wyl/BRICE.

Introduction

The introduction presents BRICE as a Predictive Coding–based recurrent encoder that uses feedforward and feedback connections to infer missing point-cloud structure. It motivates this design by limitations in existing completion architectures and by brain-inspired accounts of visual completion.

Background and challenges

  • Point-cloud completion recovers complete structure from sparse or incomplete data, whose gaps can arise from occlusion, sensor-resolution limits, or down-sampling.Completed point clouds can support downstream classification, segmentation, and 3D object detection.Source: Introduction · p. 1
  • Existing completion methods include attention-based encoder-decoders, local point-wise interaction methods, and heuristic or theory-driven approaches, yet challenges in interpretability, theoretical guidance, and feedback remain.The introduction says attention-module stacking often relies on empirical trial and error, while most studies propagate information only feedforward.Source: Introduction · pp. 1–2

BRICE motivation and design

  • BRICE uses a hierarchical Predictive Coding model in an encoder with feedforward processing and feedback guidance from deeper layers to support completion.The authors propose this recurrent design based on the idea that mechanisms associated with illusory contours in visual perception may help infer missing 3D structure.Source: Introduction · p. 2

BRICE contributions

  • BRICE is plug-and-play: replacing existing completion-network encoders with BRICE can improve multiple mainstream baselines on datasets of man-made objects and natural crops.Source: Introduction · p. 2
  • Transferring classification-pretrained BRICE to point-cloud completion can further improve multiple baselines, suggesting learned 3D shape regularities provide predictive priors for missing parts.Source: Introduction · p. 2

Related Work

Prior point cloud completion work has advanced from coarse-to-fine encoder-decoder designs to attention, cross-scale interactions, and heuristic or theoretical methods, but current models incur heavy attention computation and neglect recurrent feedback. BRICE draws on brain-inspired recurrent illusory-contour models to introduce a plug-and-play encoder for 3D completion networks.

Point Cloud Completion

  • PCN (Yuan et al. 2018) and TopNet established coarse-to-fine, two-stage completion, while global max-pooling lost fine-grained features.Later attention-based methods included PointTr (Yu et al. 2021), which used point-cloud feature sequences and a Transformer decoder, and SnowflakeNet (Xiang et al. 2021), which introduced a Skip-Transformer.Source: Point Cloud Completion · p. 2
  • Hierarchical feature interactions model spatial continuity and structural consistency; SeedFormer (Zhou et al. 2022) uses Patch Seeds, while CRA-PCN uses cross-resolution Transformer interactions.Patch Seeds capture global topology across layers while preserving local detail; CRA-PCN enables intra- and inter-layer interactions.Source: Point Cloud Completion · p. 2
  • Interpretability-oriented approaches include SymmCompletion (Yan et al. 2025), which uses local symmetry, DANCE (Kim and Cho 2026), which samples multi-view rays, and 3DDM, a physically based anisotropic diffusion model.The passage also mentions theoretical frameworks.Source: Point Cloud Completion · pp. 2–3
  • Current completion models require heavy computation of attention matrices across deep layers and neglect direct completion through recurrent loops with feedback from deeper layers.The paper frames this as the gap motivating recurrent feedback supervision.Source: Point Cloud Completion · p. 3

Brain’s Response to Illusory Contours

  • Brain studies found mammalian responses to illusory contours in low-level regions depend on higher-level feedback, inspiring recurrent feedforward-feedback models for image completion and classification.Predictive Coding (Pang et al. 2021) completed missing contours; PREDIFY (Choksi et al. 2021) applied a PC-like model to VGG for image reconstruction and denoising; DPCN (Wen et al. 2018) used the architecture for image classification and achieved comparable CIFAR10/100 performance with 7 layers against ultra-deep networks.Source: Brain’s Response to Illusory Contours · p. 3

Method

BRICE is a plug-and-play, three-layer recurrent encoder that combines feedforward processing with feedback to produce features for point-cloud completion. Its Predictive Coding updates iteratively refine representations, and the method also pretrains the encoder on point-cloud classification before transplanting it into completion networks.

Architecture of BRICE

  • BRICE replaces the encoder in mainstream completion networks with three layers that combine feedforward and feedback pathways, sending deepest-layer features to the existing decoder.The module can also be pretrained on another task and transplanted into a completion network.Source: Architecture of BRICE · p. 3

Predictive coding for updating

  • Following Predictive Coding, BRICE iteratively updates representations through recurrent self-prediction without adding physical network layers.The deepest-layer features are then processed by the completion decoder.Source: Architecture of BRICE · p. 3, Predictive coding for updating · p. 4
  • Feedforward processing maps the incomplete cloud and successive layer features through PointNet++-style set abstraction and a Transformer (Qi et al. 2017).The feedforward signal is integrated into the current layer as a feature-fusion input.Source: Predictive coding for updating · p. 4
  • A feedback sub-decoder predicts current-layer features from the next layer using MLP, deconvolution upsampling, and MLP components; L1-CD error gradients guide correction (Fan, Su, and Guibas 2017).The method’s stated hypothesis is that minimizing layer prediction errors encourages shallow features to become better predicted by higher-level features.Source: Predictive coding for updating · p. 4

Training details and losses

  • BRICE is pretrained for point-cloud classification on LHD 3D, then transplanted into completion networks; its internal objective sums reconstruction errors across recurrent steps.The completion decoder retains its original decoder and uses L1-CD between completed and ground-truth point clouds; the overall loss sums the two sub-losses.Source: Training details and losses · pp. 4–5

Experiments

BRICE improves mainstream point-cloud completion baselines across man-made and natural-object datasets, with benefits depending on dataset and decoder priors. Experiments also find that feedback, reconstruction-error correction, pretraining, and a recurrent time of 3 contribute to performance.

Datasets and training details

  • Experiments use LHD for pretraining and classification, MVP2K for L2-CD completion, and PCN and PlantPCom for L1-CD completion.The datasets contain man-made objects in LHD, MVP2K, and PCN, while PlantPCom tests generalization to natural plants.Source: Datasets and training details · p. 5

Results

  • On LHD classification, BRICE achieves 78.11% Top-1 Accuracy, outperforming the second-best CRA-PCN baseline by 0.7%.A fully connected head was appended to the BRICE encoder for this evaluation.Source: Results · p. 5, Table 1 · p. 5
  • On MVP2K, BRICE improves average L2-CD for all seven baselines, and pretraining improves all seven further; PointAttN falls from 6.95 to 5.55 (×10^-4).The comparison includes both BRICE trained from scratch and BRICE pretrained on LHD.Source: Table 2 · p. 5, Results · pp. 5–6
  • On PlantPCom, inserting BRICE improves average L1-CD across all seven baselines, with ODGNet improving by 1.02×10^-3 and becoming the best performer.BRICE is used without pretraining on this natural-plant dataset; qualitative examples show finer plant-organ details.Source: Table 4 · p. 6, Results · p. 7
  • On PCN, BRICE improves L1-CD for SnowflakeNet, SeedFormer, AnchorFormer, PointAttN, and CRA-PCN, while ODGNet and SymmCompletion decline.The authors suggest BRICE may disrupt ODGNet’s category-shared morphological prior and SymmCompletion’s symmetry prior, especially given PCN’s eight artificial-object classes.Source: Results · pp. 6–7, Table 3 · p. 6

Ablation Studies

  • Removing BRICE’s main feedback pathway reduces average L2-CD to the plain-baseline level, while removing reconstruction-error correction worsens it to 5.48×10^-4, below the plain model.The ablations use CRA-PCN on MVP2K; the authors interpret these results as evidence that deep-feature supervision and reconstruction errors support completion and dynamic updates.Source: Ablation Studies · p. 7, Table 5 · p. 7
  • Pretraining BRICE on LHD yields better completion performance than training it from scratch in the CRA-PCN ablation.The authors attribute the benefit to cross-task knowledge for completion on the same type of objects.Source: Ablation Studies · p. 7, Table 5 · p. 7
  • For CRA-PCN+BRICE+Pre on MVP2K, performance peaks at recurrent time T = 3, so the paper uses T = 3 throughout its experiments.Performance first improves and then declines as recurrent time increases.Source: Ablation Studies · p. 7, Table 6 · p. 7

Conclusion · Appendix to BRICE: Brain-inspired Recurrent Illusory Contour Encoder for Point Cloud Completion

BRICE is a brain-inspired, plug-and-play encoder for point cloud completion, but its demonstrated scope is limited to baselines with clear encoder-decoder architectures; broader applicability and future extensions remain open.

Conclusion

  • Inspired by the IC mechanism in the brain’s V1L2/3 region, BRICE is proposed as a plug-and-play encoder for enhancing point cloud completion.Source: Conclusion · p. 7
  • BRICE can replace the original encoder in various mainstream point cloud completion networks to improve performance.Source: Conclusion · p. 7
  • BRICE is currently well suited only to baselines with clear encoder-decoder architectures.Source: Conclusion · p. 7
  • Its applicability to networks without clear hierarchy or with irregularly stacked layers remains unexplored.Source: Conclusion · p. 7
  • Future work will reduce BRICE’s parameters, test it on more diverse 3D datasets, and design versions for unified vision tasks.Source: Conclusion · p. 7

A. The Details of the datasets

The experiments use LHD to pre-train BRICE for classification and MVP2K, PCN, and PlantPCom to evaluate point-cloud completion. The datasets differ in object domains, splits, incomplete-view generation, and evaluation metrics.

A. The Details of the datasets

  • LHD contains 235,175 point clouds across 87 categories and is used to pre-train BRICE for classification, evaluated by Top-1 Accuracy (Chen et al. 2023a).It has 183,704 training and 51,471 validation examples; point clouds are resized to 2,048 points, and 99.8% are man-made objects.Source: A. The Details of the datasets · p. 10
  • MVP2K comprises 100,000 point clouds from 16 categories, with 62,400 for training and the rest for testing; completion is evaluated using L2 Chamfer Distance (L2-CD) (Pan et al. 2021).Each object is rendered from 26 spherical camera viewpoints, producing incomplete views; incomplete and ground-truth clouds are fixed to 2,048 points.Source: A. The Details of the datasets · p. 10
  • PCN contains 30,974 ShapeNet point clouds from 8 man-made categories, split into 28,974 training, 800 validation, and 1,200 test examples, and uses L1-CD for evaluation (Yuan et al. 2018).Input point clouds are fixed at 2,048 points.Source: A. The Details of the datasets · p. 10
  • PlantPCom contains 20,800 plants from 5 species, uses an 8:1:1 training-validation-test split, and evaluates completion with L1-CD (Li et al. 2025).Its incomplete clouds are generated from high-precision plant models across 20 viewpoints, and each input cloud is fixed at 2,048 points.Source: A. The Details of the datasets · p. 10

B. The training details

The section specifies training protocols for the datasets and pre-training of BRICE, including optimizer, learning-rate, batch-size, and model-selection choices where reported.

B. The training details

  • BRICE pre-training followed the configuration suggested in PointGPT (Chen et al. 2023a).Source: B. The training details · p. 10
  • MVP2K and PlantPCom baselines used Adam with weight decay 0.001; learning rates were 0.0001 and 0.001, and batch sizes were 48 and 32, respectively.Source: B. The training details · p. 10
  • Both MVP2K and PlantPCom training used learning-rate warmup, ran for 100 epochs, and selected the model with the lowest training loss for testing.Source: B. The training details · p. 10
  • PCN baselines used the recommended hyperparameter settings from their respective publications (Xiang et al. 2021; Zhou et al. 2022; Chen et al. 2023b; Wang et al. 2024; Cai et al. 2024; Rong et al. 2024; Yan et al. 2025).Source: B. The training details · p. 10

C. Efficiency of learning · D. Detailed architectures of evaluated baselines/networks

On MVP2K, evaluated baselines with and without BRICE took around 5.4 to 30.2 hours to train, while BRICE-enhanced baselines had inference times of 9.13 to 74.14 ms per point cloud. BRICE was also integrated into eight mainstream completion networks, with original and integrated architectures visually compared.

C. Efficiency of learning

  • On MVP2K, evaluated baselines, both with and without BRICE, took around 5.4 to 30.2 hours to finish training; BRICE-enhanced baselines averaged 9.13 to 74.14 ms inference time per single point cloud.Source: C. Efficiency of learning · p. 11
  • 9.13 to 74.14 ms was the average inference time per point cloud for BRICE-enhanced baselines.The passage describes this range as applying to a single point cloud.Source: C. Efficiency of learning · p. 11
  • BRICE's parameter count is mainly due to its up-sampling and three transformer blocks.Source: C. Efficiency of learning · p. 11

D. Detailed architectures of evaluated baselines/networks

E. More quantitative results on MVP2K dataset

This section presents full quantitative results for all MVP2K categories in Table A2, extending the manuscript’s space-limited presentation of five categories and the dataset average.

  • Table A2 lists quantitative MVP2K results for all categories; the manuscript previously showed only Airplane, Lamp, Sofa, Table, Watercraft, and the total average.The full category results are provided to comprehensively demonstrate BRICE’s effectiveness.Source: E. More quantitative results on MVP2K… · p. 12

F. Qualitative results on PCN dataset

On PCN qualitative examples, BRICE improves completion with smoother surfaces and more realistic point distributions.

F. Qualitative results on PCN dataset

  • BRICE smooths structural gaps or rough surfaces in the Couch top and Boat mast completions of SeedFormer and CRA-PCN, producing more realistic point distributions.These qualitative examples are shown in Fig. A2.Source: F. Qualitative results on PCN dataset · p. 13

G. Qualitative results on PlantPCom Dataset

On PlantPCom qualitative examples, BRICE improves SeedFormer’s stem morphology and suppresses noise, while removing separated noise clusters produced by PointAttN.

G. Qualitative results on PlantPCom Dataset

Related papers

Loading 2610.23001v1…