acceptodds
Under review as a conference paper at ICLR 2027

Brain-inspired Recurrent Illusory Contour Encoder for Point Cloud Completion

Abstract

Point cloud completion refers to the process of recovering a complete structural representation from incomplete or sparse 3D point cloud data. Currently, mainstream point cloud completion methods are unanimously based on deep learning architectures. On one hand, these methods tend to repetitively stacking complex attention modules to enhance network performance, which drifts away from the neat design of unified and interpretable networks; on the other hand, nearly all existing architectures rely solely on feedforward connections to propagate information, which lacks feedback and supervision from deeper layers. The recurrent neural architecture, formed by feedforward-feedback loops, have shown its ability to generate Illusory Contours (ICs) in the V1 region of brains from several species of mammals; and ICs (as exemplified by the Kanizsa triangle) are essentially a completion mechanism of the brain for visual information. Inspired by this observation, 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 it to be seamlessly integrated into mainstream point cloud completion networks. Experimental results have shown that BRICE can enhance various point cloud completion baselines on 3D datasets such as MVP2K, PCN, and PlantPCom.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Point cloud completion, Illusory contours, Predictive coding, Brian-inspired recurrent encoder

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_sP9mOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes BRICE, a brain-inspired recurrent encoder for point cloud completion. By incorporating feedforward-feedback connections under the Predictive Coding framework, BRICE aims to improve feature representations and enhance the recovery of missing geometric structures. The proposed module is designed to be plug-and-play and is evaluated across multiple point cloud completion baselines and benchmarks. While the brain-inspired perspective is interesting and the experiments demonstrate certain improvements, I have concerns regarding the methodological novelty and the consistency of performance gains.

Critical strengths
  • This paper proposes a plug-and-play recurrent module for point cloud completion and evaluates its effectiveness across multiple baselines and benchmarks.

  • The brain-inspired perspective on improving point cloud completion is interesting. The extensive quantitative comparisons, qualitative visualizations, and ablation studies provide some evidence supporting the effectiveness of the proposed method.

Critical weaknesses
  • I have some concerns regarding the originality of the proposed method. In particular, the differences between BRICE and the existing work FBNet [1] should be further clarified. Both methods explore feedback mechanisms for point cloud completion. The authors are encouraged to provide a more detailed comparison and explicitly highlight the methodological contributions of BRICE beyond existing feedback-based approaches.

  • Although BRICE improves the performance of several baselines, the improvements are relatively marginal in some cases. For example, in Table 2, the improvement over the original SnowflakeNet is almost negligible. Moreover, as shown in Table 3, integrating BRICE into SymmCompletion and ODGNet even leads to performance degradation. These results raise concerns about the consistency and general applicability of the proposed module. Further discussion of the conditions under which BRICE is effective would strengthen the experimental analysis.

[1] FBNet: Feedback Network for Point Cloud Completion, ECCV 2022.

Questions for authors

See Critical Weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer sP9m (1/2)

Dear Reviewer,

Thank you very much for carefully reviewing our work and for your valuable feedback. We have now uploaded the revised manuscript. In the PDF file, all newly added or modified sections have been highlighted in yellow for easier identification. Additionally, to provide a deeper understanding of our work, we have included an explanatory video in the supplementary materials. We kindly invite you to review it at your convenience.

W1: I have some concerns regarding the originality of the proposed method. In particular, the differences between BRICE and the existing work FBNet [1] should be further clarified. Both methods explore feedback mechanisms for point cloud completion. The authors are encouraged to provide a more detailed comparison and explicitly highlight the methodological contributions of BRICE beyond existing feedback-based approaches.

Response: We fully agree that FBNet (ECCV 2022) [1] is an excellent and pioneering work exploring feedback mechanisms for point cloud completion. However, BRICE differs from FBNet in several fundamental aspects. First, they differ in the location of the feedback (recurrent structure). BRICE's recurrence occurs at the very front end of the network, whereas FBNet employs recurrence in the later stages, specifically within the decoder. Second, the computational mechanisms are different. BRICE utilizes a recurrent computational mechanism guided by predictive coding (modeling the primary visual cortex), featuring layer-wise feature reconstruction and feedback. In contrast, FBNet relies on standard implicit multi-layer feature feedback without a strong reconstruction process. Third, BRICE is proposed as a universal, plug-and-play encoder capable of enhancing most mainstream point cloud completion networks, whereas the feedback recurrent computation in FBNet does not share this general applicability. Fourth, the completion performance of BRICE combined with mainstream networks significantly surpasses that of FBNet.

We even conducted an interesting experiment by attempting to enhance FBNet using BRICE. We evaluated the direct integration of BRICE and FBNet on the MVP2K dataset. The results are as follows:

Table: Experiments for FBNet on the MVP2K by L2-CD. Omitted by ×10−4\times 10^{-4} for readability.

BaselineAvg. ↓\downarrowParams(M)Flops (G)
FBNet6.313.8210.45
+BRICE6.5314.6526.73
CRA-PCN5.465.684.63
+BRICE5.3517.5120.03
SymmCompletion5.34‾\underline{5.34}6.647.97
+BRICE5.07\mathbf{5.07}19.3624.67

As shown in the table above, forcibly combining BRICE (a dynamic encoder with feedback) and FBNet (an iterative decoder with feedback) not only fails to improve performance (the error increases from 6.31 to 6.53) but also causes a severe surge in network parameters and computational overhead. This failure stems from conflicts in their neurobiological plausibility and underlying structural designs. A detailed analysis is as follows: First, the BRICE+FBNet hybrid forms a full-chain, physical-level loop. This forces every layer of the network to attempt to reconstruct the point cloud, causing the "incomplete" physical morphology of the input to be propagated all the way to the end, which is actually detrimental to the recovery of missing high-frequency information. Second, the feedback mechanisms in BRICE and FBNet are fundamentally different, leading to a functional mismatch upon integration. FBNet’s feedback is implicit and macroscopic (re-encoding the output point cloud and concatenating it with previous layer features). In contrast, BRICE is explicit and microscopic, directly computing the Chamfer Distance (CD) reconstruction error between the top-down predicted features and the bottom-up actual features within the latent feature space.

AuthorsOpenReview ↗

Response to Reviewer sP9m (2/2)

W2: Although BRICE improves the performance of several baselines, the improvements are relatively marginal in some cases. For example, in Table 2, the improvement over the original SnowflakeNet is almost negligible. Moreover, as shown in Table 3, integrating BRICE into SymmCompletion and ODGNet even leads to performance degradation. These results raise concerns about the consistency and general applicability of the proposed module. Further discussion of the conditions under which BRICE is effective would strengthen the experimental analysis.

Response: As you have noted, because BRICE is a plug-and-play module built upon explicit theoretical priors, its effectiveness indeed possesses specific applicability boundaries. We candidly attribute these phenomena to the following factors, which also represent the limitations of our current work: SnowflakeNet intrinsically incorporates an exceptionally powerful and complex 'Snowflake Point Deconvolution (SPD)' mechanism. This decoder has already pushed the extraction of local geometric features to an extreme engineering limit. In such scenarios, the performance bottleneck of the overall network is often no longer the feature representation capacity of the encoder, causing the gains introduced by BRICE to appear relatively marginal. However, even on such a formidable baseline, BRICE still ensures a stable performance lower bound without requiring any alterations to the decoder's physical architecture.

A detailed explanation regarding the performance degradation observed with SymmCompletion and ODGNet, along with a discussion of BRICE's optimal operating conditions, has been provided in Section 4.2 of the manuscript. The specific details are as follows:

On the PCN dataset, the undesirable adaptation of BRICE on ODGNet may relate to the dictionary-guided decoder of ODGNet that strongly abstracts common morphological features from all objects of the same category as a strong prior knowledge to facilitate 3D completion. This decoder design is well-suited for man-made objects (such as planes, vehicles, etc.) from which shared structural knowledge can be easily generalized, and the limited diversity of objects (only 8 classes) of PCN further amplifies the effect of the decoder. Adding BRICE to ODGNet probably breaks the formation of shared knowledge. BRICE’s failing to improve SymmCompletion may due to the latter's tailored decoder for 3D geometric symmetry as prior knowledge. Given that all 8 object classes in the PCN dataset are artificial objects with strong bilateral symmetry, the integration of BRICE disrupts the symmetry prior, leading to a decline in performance.

[1] Xuejun Yan, Hongyu Yan, Jingjing Wang, Hang Du, Zhihong Wu, Di Xie, Shiliang Pu, and Li Lu. Fbnet: Feedback network for point cloud completion. In European conference on computer vision, 2022.

We hope our responses have successfully addressed your concerns and clarified any questions you had regarding the manuscript. Please let us know if you require any further discussion or clarification!

Thank you to the authors for their response and additional experiments. After considering the other reviewers’ comments, I remain concerned that the largest reported gain relies on a PointAttN baseline that the authors acknowledge did not converge optimally. The gains are also inconsistent across baselines, and the new controls do not match computational cost. These issues limit the strength of the paper’s general claims. I therefore maintain my original score.

I encourage the authors to continue refining the paper. With a stronger experimental evaluation and more carefully scoped claims, I believe this work can find a suitable venue.

AuthorsOpenReview ↗

Response to Reviewer sP9m

Dear Reviewer,

Thank you for your follow-up. We would like to clarify the misunderstanding caused by the improper phrasing in our previous response:

We must sincerely apologize for the inaccurate statement ("failed to reach an optimal convergence state") in our previous rebuttal, which contradicted our actual experimental results and misled your judgment. First, PointAttN (6.95) fully reached optimal convergence rather than failing to converge; we have uploaded the training loss curves of PointAttN, +BRICE, and +BRICE+pre in the attachment, which clearly demonstrate that it has reached convergence. The reason for our poor phrasing in the previous round is that the original PointAttN paper did not report results on MVP2K; what we actually intended to convey was that "PointAttN’s native architecture suffers from an inherent representation bottleneck under the multi-category (16 classes) and sparser input setting of MVP2K, resulting in a slightly higher absolute error (6.95) compared to more recent baselines such as CRA-PCN," rather than suggesting that it was not trained to convergence.

Furthermore, our four newly added independent control experiments directly verify this: the original PointAttN (6.95), w/ equal-budget feedforward with added feedforward layers (6.99), w/ generic recurrent with standard recurrence (6.88), and w/ pre with LHD pretraining (6.87) all stably converged within a very narrow range of 6.87–6.99. This demonstrates that around 6.95 reflects the true performance of the PointAttN encoder after full convergence on MVP2K, and it is only through the introduction of BRICE that this representation bottleneck is broken (6.95→5.62, and 5.55 with pretraining).

We sincerely hope that these clarifications and the attached training curves address your remaining concerns, and we would be deeply grateful if you could reconsider your evaluation of our work.

Best regards,

Authors

Reviewer_8QzhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a recurrent point-cloud encoder that combines bottom-up features, top-down predictions, and reconstruction-error gradients to refine representations before completion. Its main contribution is adapting predictive-coding dynamics into a replacement encoder for existing completion networks. Experiments cover seven backbones on MVP2K, PCN, and PlantPCom, with optional classification pretraining on LHD. The reported improvements are accompanied by increased computation and are not consistent across all backbones, particularly on PCN.

Critical strengths
  • BRICE adapts top-down prediction and reconstruction-error correction into a recurrent point-cloud encoder. This is a useful application of established predictive-coding principles, with potential to improve representations for completion.

  • Revisiting low-level features using higher-level context is a plausible way to improve completion under occlusion. So The proposed internal prediction mechanism is more specific than simply adding another attention block.

  • On MVP2K, PointAttN improves from 6.95 to 5.55 in the reported L2-CD units with pretrained BRICE. On PlantPCom, ODGNet improves from 5.11 to 4.09 in the reported L1-CD units.

Critical weaknesses
  • BRICE receives additional supervised classification training on LHD, but the original encoders are not evaluated with equivalent pretraining. The +BRICE versus +BRICE+Pre comparison establishes the benefit of pretraining BRICE, but does not show whether this benefit is specific to its design or also attainable with the baseline encoders.

  • The neuroscience discussion provides motivation, but the reported experiments measure geometric completion error. They do not demonstrate illusory-contour responses in the model or show that its mechanism corresponds to the cited cortical mechanism. Similarly, the optimum at three recurrent steps does not establish an analogy with the temporal disappearance of visual illusions; ordinary optimization or oversmoothing effects could explain the same observation.

  • Table 7 shows substantial parameter increases: SeedFormer grows from 2.64M to 17.54M and CRA-PCN from 5.68M to 17.51M. Without parameter- and compute-matched controls, it remains unclear how much improvement comes from predictive coding rather than greater capacity and computation.

Questions for authors
  1. How does BRICE compare with equal-budget feedforward, generic recurrent, and auxiliary-reconstruction-only encoders?
  2. What exact computational graph implements G_l, and are higher-order derivatives retained during training?
  3. Can you give the original encoders the same LHD classification pretraining, with matched data and training budgets, and compare their subsequent completion performance with +BRICE+Pre?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 8Qzh (1/2)

Dear Reviewer,

Thank you very much for carefully reviewing our work and for your valuable feedback. We have now uploaded the revised manuscript. In the PDF file, all newly added or modified sections have been highlighted in yellow for easier identification. Additionally, to provide a deeper understanding of our work, we have included an explanatory video in the supplementary materials. We kindly invite you to review it at your convenience.

Q1: How does BRICE compare with equal-budget feedforward, generic recurrent, and auxiliary-reconstruction-only encoders?

Response: Your suggestion is highly constructive. Using the PointAttN baseline network as an example, we conducted detailed comparative experiments, as shown in the table below. First, regarding the 'Equal-parameter-budget' setup: when we increased the depth of the native feedforward layers to match the parameter count of BRICE, the performance actually degraded compared to the plain PointAttN (6.95→6.99). However, at the exact same parameter budget of approximately 21.00M, integrating BRICE yielded a significant performance improvement (6.95→5.62). This demonstrates that the core source of the performance gain is the dynamic recurrent mechanism of predictive coding, rather than a mere expansion of parameter capacity.

Second, regarding the 'generic recurrent' baseline, we constructed a simple recurrent control group by introducing weight-shared time-step recurrent iterations during the deep feature extraction stage (also setting T=3T=3). Our experiments indicate that while this purely repetitive feature recurrence increases the non-linear depth of the network, it easily leads to feature over-smoothing during the iterations. Consequently, its final completion accuracy falls short of the full BRICE model (6.88 vs. 5.62). This further confirms that a simple unrolling of time-step recurrence is not the key factor. Rather, the essential reason BRICE breaks through the performance bottleneck lies in its ability to dynamically and directionally reshape the feature manifold.

Finally, regarding the 'auxiliary-reconstruction-only' baseline, we have provided a clear answer in the ablation study in Table 5 of the main text via the 'w/o Recons' variant. By removing the dynamic gradient correction of features based on the reconstruction error within the recurrent loop, the network essentially degrades into a standard feedforward encoder with merely an auxiliary reconstruction objective during training. As shown in Table 5, the L2-CD error of this variant increases to 5.48, performing even worse than the plain model (5.46). In contrast, the full model, which retains the complete error correction feedback mechanism, achieves the optimal performance of 5.29.

Table: Ablation experiments for PointAttN on the MVP2K by L2-CD. Omitted by ×10−4\times 10^{-4} for readability.

Model variantsAvg. ↓\downarrowParams(M)Flops (G)
PointAttN6.9519.006.17
w/ equal-budget feedforward6.9921.196.69
w/ generic recurrent6.8821.197.73
w/ pre6.8719.006.17
+BRICE5.62‾\underline{5.62}21.0022.09
+BRICE+pre5.55\mathbf{5.55}21.0022.09

Q2: What exact computational graph implements G_l, and are higher-order derivatives retained during training?

Response: GlG_l is computed using PyTorch’s standard Autograd engine. In our actual training implementation, we do not retain higher-order derivatives. To prevent the memory explosion caused by unrolling across time steps, once GlG_l is obtained, we implicitly truncate the computational graph (e.g., via the .detach() operation). Consequently, GlG_l serves purely as a directional signal fed into the feature reshaping process of the subsequent time step. This approach maintains training efficiency while ensuring the effectiveness of the mechanism.

AuthorsOpenReview ↗

Response to Reviewer 8Qzh (2/2)

Q3: Can you give the original encoders the same LHD classification pretraining, with matched data and training budgets, and compare their subsequent completion performance with +BRICE+Pre?

Response: Why introduce pre-training (+BRICE+pre)? The introduction of pre-training was not merely intended to boost performance; rather, it was driven by the theoretical requirements of our 'brain-inspired' methodology. Studies indicate that human capabilities for contour integration and grouping are closely tied to the statistical regularities of the natural world [1]. To equip the model with sufficient 'contour prior knowledge' to successfully drive the top-down predictive coding mechanism, pre-training on a large-scale dataset (LHD) is conceptually indispensable.

However, to address your question more clearly, we selected the representative baseline network PointAttN to conduct strictly controlled experiments, as shown in the table above. Maintaining the exact same data scale and training configurations, we applied identical classification pre-training to the original PointAttN encoder on the LHD dataset (denoted as PointAttN w/ pre). The classification accuracy of this pre-training (68.97%) is also reported in Table 1 of the main text. Our results demonstrate that although pre-training on LHD yields some performance improvements for the original baseline, its final completion accuracy remains significantly inferior to our +BRICE+pre (6.87 vs. 5.55). Therefore, BRICE's performance advantage does not merely stem from 'seeing more pre-training data.' On the contrary, it is precisely because of BRICE's unique predictive coding and dynamic error-correction mechanisms that the network can effectively translate the 'natural contour priors' learned during the pre-training phase into a powerful guiding force for reconstructing missing geometric structures.

[1] Geisler, W. S., & Perry, J. S. Contour statistics in natural images: Grouping across occlusions. Visual Neuroscience, 2009.

We hope our responses have successfully addressed your concerns and clarified any questions you had regarding the manuscript. Please let us know if you require any further discussion or clarification!

Reviewer_DkWnOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes BRICE, a plug-and-play recurrent encoder for point cloud completion motivated by illusory-contour processing in V1. A three-layer encoder (set abstraction + Transformer per layer, Fig. 2c) is unrolled T=3 times under a predictive-coding update (Eq. 1) combining a feedforward term, a feedback prediction from an MLP+deconvolution sub-decoder (Eq. 3), a memory term, and the gradient of an "L1 Chamfer" error computed on intermediate feature sets (Eq. 4). Summed inter-layer errors form an auxiliary loss L_BRICE (Eq. 6) added to the host decoder's loss. BRICE replaces the encoders of seven completion baselines; the paper reports improved average CD for all seven on MVP2K (Table 2) and PlantPCom (Table 4) and five of seven on PCN (Table 3), discloses the two PCN failures with analysis, tabulates the 3–9× parameter and 3–5× FLOP overhead (Table 7), and ablates feedback, error correction, pre-training, and T on CRA-PCN/MVP2K (Tables 5–6).

Critical strengths

The ablation suite contains the right capacity control, and it favors the authors. For a module adding 2–19M parameters, capacity is the decisive confound, and Tables 5–6 address it: "w/o Main feedback" (λ=0) scores 5.46, exactly plain CRA-PCN's 5.46, and T=0 through the enlarged BRICE trunk also returns exactly 5.46, while the full model reaches 5.29. Taken at face value, the extra ~12M parameters alone buy nothing on this pair and the gain is attributable to the recurrent dynamics. A corroborating pattern: the largest MVP2K gain (PointAttN, 6.95 → 5.62) coincides with the smallest parameter delta (+2.0M, Table 7), arguing against a pure parameter-count explanation. The control exists for one baseline on one dataset, run once — but it is the right experiment.

Honest disclosure of failures and costs. The paper reports the two PCN baselines BRICE harms (ODGNet 6.49 → 6.81, SymmCompletion 6.33 → 6.48) with discussion rather than concealment, fully tabulates overhead (SeedFormer 2.64M → 17.54M parameters, ODGNet 4.48M → 23.92M, inference up to ~2.9×), reports a non-monotonic T-sweep, and concedes a scope limitation. This transparency makes the evidence auditable — and auditing it is exactly what reveals the protocol problems below.

Critical weaknesses
  1. The headline gains do not survive comparison with the published literature, and the quantitative record is unverifiable from the paper itself. On MVP2K and PlantPCom, all baselines were retrained under one uniform recipe — 100 epochs, shared optimizer/learning rate/batch size, checkpoint selected by lowest training loss — while PlantPCom's existing 8:1:1 validation split goes unused. This recipe demonstrably weakens baselines: CRA-PCN's own paper reports 5.33 on MVP-2048 versus the 5.46 reproduced here, so the best +BRICE+Pre result (5.29) is only 0.04 better than the published plain baseline; and PointAttN's 6.95 is far out of family with every other retrained baseline (5.34–6.48) while supplying the paper's largest advertised win. On PCN — the one benchmark where each method's own recipe was used — five of seven reproductions are weaker than published (SnowflakeNet 7.40 vs 7.21, AnchorFormer 6.79 vs 6.59, PointAttN 6.93 vs 6.86, CRA-PCN 6.56 vs 6.39, SymmCompletion 6.33 vs 6.28). Against published numbers, baseline+BRICE wins only twice: SeedFormer (6.46 vs 6.74) and marginally PointAttN (6.83 vs 6.86); SnowflakeNet+BRICE (7.25), AnchorFormer+BRICE (6.72), CRA-PCN+BRICE (6.42) and SymmCompletion+BRICE (6.45) all remain worse than the plain published baselines they are meant to improve. Worse, the averages themselves cannot be checked: no printed Avg equals the mean of its printed row in Tables 2, 4, or 8 (e.g., SnowflakeNet's 16-category mean is 5.74 against a printed 6.05; the deviations are systematic and consistent with undisclosed per-sample weighting, which is nowhere stated); Table 3 prints only 5 of PCN's 8 categories, so its Avg column is unverifiable; Tables 2 and 8 disagree on SymmCompletion+BRICE (5.07 vs 5.06); SymmCompletion's two PCN rows print identical values on 3 of 5 categories across two different training modes; and no seeds, variance, or significance tests appear anywhere, although the "all seven baselines" claim on MVP2K rests partly on SnowflakeNet's 6.05 → 6.05 → 6.04.

  2. The method as written cannot be re-implemented, and both the framing and the update rule understate prior work. (i) Eq. 7 is a verbatim duplicate of Eq. 6, confirmed in the rendered PDF: L_Decoder, half of L_Total, is never defined, and the one-line gloss does not match the multi-stage losses the seven host decoders actually use. (ii) The novel ingredient — Chamfer distance on intermediate features — is undefined: Eq. 5 defines L1-CD over point sets (with self-contradictory "squared L1" norm notation), yet Eq. 4/6 apply it to feature matrices; what the "points" are, the normalization, and the degenerate minimum obtained by shrinking feature norms are all unaddressed. (iii) Indices are inconsistent: the text defines the layer-l error as L_CD(E_l, D_l) while Eq. 4 differentiates the (l−1)-level error, and Eq. 6's sum to l=3 includes a term requiring a D_3 that cannot exist, since Eq. 3 generates D_{l−1} from E_l and there is no fourth layer. (iv) Eq. 1 has the same four-term structure as the Predify update of Choksi et al. (2021) — cited in the paper — and Appendix B adopts β=0.8, λ=0.1, α=0.01, which are exactly the coefficient values Predify's main experiments set; the equivalence is never acknowledged, leaving the update rule's novelty overstated. (v) The framing "nearly all existing architectures rely solely on feedforward connections" and "for the first time, we design a plug-and-play recurrent encoder" is contradicted by FBNet — "Feedback Network for Point Cloud Completion," ECCV 2022 oral — present in the paper's own reference list, cited only inside citation clusters, and never discussed or compared; its released model reports 5.05 CD-T on the same MVP-2048 test set, equal to the best number anywhere in this submission (SymmCompletion+BRICE+Pre, 5.05).

Compounding presentation defects: Figures 4, 5 and 6 all carry the same wrong caption; Appendix E says Table 2 shows five categories when it shows eight; ODGNet's PlantPCom gain is printed as 1.02×10⁻⁴ in a table whose units are ×10⁻³; the optimizer is specified as "β1 = 0.9 and β1 = 0.999"; and Rao & Ballard (1999), cited in Section 3.2 as the foundation of the method, is missing from the reference list.

Questions for authors

Some clarifications:

  1. Correct Eq. 7 and state L_Decoder exactly for each host network: did every baseline keep its original multi-stage decoder loss with L_BRICE added, or was any replaced by a single output-vs-GT CD? Define L_CD on features precisely (which sets, what dimensionality, what normalization; how is trivial error reduction by shrinking feature norms prevented), and state what the l=3 term of Eq. 6 is, given that no D_3 exists.
  2. Justify checkpoint selection by lowest training loss with no validation; report each plain baseline under its own published recipe on MVP2K (CRA-PCN's own paper reports 5.33 against your 5.46); explain the out-of-family PointAttN 6.95 row; and provide mean±std over at least 3 seeds for Table 5 and for the smallest-margin Table 2 rows.
  3. Define the "Avg." column: no printed average equals the mean of its printed row in Tables 2, 4, or 8; if it is sample-weighted, state the per-category test counts. Resolve the 5.07/5.06 discrepancy between Tables 2 and 8, explain the duplicated SymmCompletion PCN values across two training modes, and print the three missing PCN categories behind Table 3's Avg.
  4. Reconcile the "nearly all feedforward" framing and "for the first time" with FBNet in your own bibliography: in what precise sense is BRICE first? Add a direct FBNet comparison on MVP, and position Eq. 1 against the Predify update given the identical structure and coefficient values.
5 comments
AuthorsOpenReview ↗

Response to Reviewer DkWn (Q1)

Dear Reviewer,

Thank you very much for carefully reviewing our work and for your valuable feedback.

First, we sincerely apologize for these minor oversights: the incorrect figure captions (Fig. 4-6), the unit error for ODGNet, the typo regarding the optimizer (β1 = 0.9 and β2 = 0.999), the description error in Appendix E (8 categories instead of 5), and the missing citation for Rao & Ballard (1999). All of these have been corrected in the revised manuscript.

In the PDF file, all newly added or modified sections have been highlighted in yellow for easier identification. Additionally, to provide a deeper understanding of our work, we have included an explanatory video in the supplementary materials. We kindly invite you to review it at your convenience.

Q1: Correct Eq. 7 and state L_Decoder exactly for each host network: did every baseline keep its original multi-stage decoder loss with L_BRICE added, or was any replaced by a single output-vs-GT CD? Define L_CD on features precisely (which sets, what dimensionality, what normalization; how is trivial error reduction by shrinking feature norms prevented), and state what the l=3 term of Eq. 6 is, given that no D_3 exists.

Response:ⅰ. Equation 7 was a typographical error resulting from copying Equation 6. In our actual implementation, all baseline networks fully retain their original multi-stage decoder losses, and the LBRICEL_{BRICE} is added to the total loss solely as an auxiliary regularization term. To improve clarity and ensure rigorous presentation, we have rewritten this section. The specific modifications are as follows:

When inserting BRICE into a completion network, we strictly preserve the original architecture and the specific training objectives of its decoder. Taking the typical multi-stage baseline as an example, LDecoderL_{Decoder} is calculated as the sum of Chamfer Distances over all subsequent coarse-to-fine outputs:

L_{Decoder}=\sum_{i=1}^n L_{CD}\left(P_i,G\right) \tag{7}

where PiP_{i} denotes the output at the ii-th coarse-to-fine stage, and GG is the corresponding ground truth (GT). nn represents the number of refinement stages.

ⅱ. Feature CD Definition and Trivial Solution Prevention: The feature CD is computed over an unordered set of feature vectors. To prevent trivial solutions (i.e., feature norms shrinking to zero), we do not detach the gradients during the forward pass. The bottom-up features ElE_{l} are strongly anchored by the primary task loss LDecoderL_{Decoder} , and the role of LBRICEL_{BRICE} is to pull the top-down predictions toward this anchored representation, rather than causing the feature space to collapse.

ⅲ. We sincerely apologize for the confusion regarding the indexing: indeed, there is no fourth layer, and therefore D3D_3 does not exist. For the top layer (l=3l=3), the feedback coefficient λ\lambda is strictly set to 0, which naturally eliminates the D3D_3 term. The revised manuscript now clearly specifies the boundary conditions for all equations. Please refer to the revised manuscript for the modified equations.

AuthorsOpenReview ↗

Response to Reviewer DkWn (Q2)

Q2: Justify checkpoint selection by lowest training loss with no validation; report each plain baseline under its own published recipe on MVP2K (CRA-PCN's own paper reports 5.33 against your 5.46); explain the out-of-family PointAttN 6.95 row; and provide mean±std over at least 3 seeds for Table 5 and for the smallest-margin Table 2 rows.

Response:ⅰ. Justification for Checkpoint Selection: Selecting the checkpoint based on the minimum training loss was not an arbitrary choice. We strictly adhered to the training and evaluation protocol provided in the official open-source codebase of CRA-PCN. Given the lack of a standardized validation split for the MVP2K dataset across various open-source projects, we uniformly adopted the official CRA-PCN configuration to ensure a fair comparison of all baseline networks under the exact same standard.

ⅱ. Regarding the performance gap for CRA-PCN (5.46 vs. 5.33): Since the codebase and evaluation protocols are identical, the marginal difference between our reproduced 5.46 and the original paper's 5.33 primarily stems from variations in the underlying hardware architecture and software environments (e.g., different GPU models, PyTorch/CUDA versions). It is notoriously difficult to reproduce the exact absolute metric values reported in original papers across different computing environments. This is precisely why we insisted on retraining all 'original baselines' and their '+BRICE' counterparts from scratch within our own unified environment, thereby ensuring absolute fairness when measuring relative improvements. Furthermore, since the other six methods did not report training results on the MVP2K dataset in their original papers, we applied the exact same training protocol used for CRA-PCN to them as well, further guaranteeing the fairness of our experiments.

ⅲ. Regarding the anomaly with PointAttN (6.95): Upon adopting this unified CRA-PCN protocol, PointAttN failed to reach an optimal convergence state (achieving 6.95). This indicates that the original PointAttN is highly dependent on its specific hyperparameter tuning or bespoke validation scheduler. However, the most crucial takeaway is that even when a baseline network converges poorly, its performance is drastically improved once integrated with BRICE (6.95→5.62). This compellingly demonstrates the strong robustness and feature error-correction capabilities of BRICE.

ⅳ. Regarding the variance experiments: We fully agree with your suggestion. We have rerun the experiments for Table 5, as well as the rows in Table 2 that showed smaller margins of improvement, using three different random seeds (1111, 1115, and 1128). We have supplemented our reporting with the mean and standard deviation (mean±std), as shown in the table below. This stable statistical data conclusively demonstrates that the performance gains genuinely stem from the model's architectural mechanisms, rather than random fluctuations.

Table: Mean-Variance experiment on the MVP2K by L2-CD. Omitted by ×10−4\times 10^{-4} for readability.

MethodsSeed (1111)Seed (1115)Seed (1128)Avg. ↓\downarrow
SnowflakeNet6.126.046.056.07±0.046.07\pm0.04
+BRICE6.096.046.056.06±0.036.06\pm0.03
+BRICE+pre6.066.026.046.04±0.026.04\pm0.02
Plain model5.485.395.465.44±0.055.44\pm0.05
w/o Main feedback5.425.435.465.44±0.025.44\pm0.02
w/o Recons5.455.505.485.48±0.035.48\pm0.03
w/o Pretraining5.385.325.355.35±0.035.35\pm0.03
Full model5.285.315.295.29±0.025.29\pm0.02
AuthorsOpenReview ↗

Response to Reviewer DkWn (Q3)

Q3: Define the "Avg." column: no printed average equals the mean of its printed row in Tables 2, 4, or 8; if it is sample-weighted, state the per-category test counts. Resolve the 5.07/5.06 discrepancy between Tables 2 and 8, explain the duplicated SymmCompletion PCN values across two training modes, and print the three missing PCN categories behind Table 3's Avg.

Response:ⅰ. Clarification on Mean Values: All mean values reported in the tables are weighted averages based on the number of samples in each category, rather than simple arithmetic means. We wish to emphasize that for every dataset, we strictly adhered to the calculation protocols established by prior works to ensure a fair comparison. Specifically, the number of test samples per category varies in the MVP dataset, whereas the PCN dataset contains a uniform number of test samples across all categories. The weighted average is calculated using the following formula:

CDavg=∑c=1C(CDc×Nc)∑c=1CNcCD_{avg}=\frac{\sum_{c=1}^C{\left( CD_c\times N_c \right)}}{\sum_{c=1}^C{N_c}}

where CC denotes the total number of categories, CDcCD_c represents the average CD for all samples in category cc, and NcN_c is the number of test samples for category cc.

ⅱ. Clarification on Data: The discrepancy (5.07 vs. 5.06) between Table 2 and Table 8 was caused by a data transcription error. We sincerely apologize for this oversight, which has now been corrected. Furthermore, we have re-verified the three identical data points for SymmCompletion on the PCN dataset and found no anomalies. To address your question more clearly, we have supplemented the main text with the missing data for the three categories in Table 3.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Overall Response to Reviewers

Dear Reviewers,

We sincerely thank all reviewers for their detailed, professional, and highly insightful evaluation of our manuscript. We are greatly encouraged by your recognition of the paper's strengths, including: the rigorous and convincing capacity-controlled ablation studies (Reviewer DkWn), the candid and transparent disclosure of limitations (Reviewer DkWn), the novel and practical brain-inspired predictive coding perspective (Reviewer 8Qzh), and the solid, extensive qualitative and quantitative validations (Reviewer sP9m).

During the rebuttal period, we have carefully incorporated every constructive suggestion raised by the reviewers. We have not only corrected the typos in the text but also added a substantial number of comparative experiments and theoretical clarifications, resulting in a comprehensive revision of the manuscript. (In the PDF file, all newly added or modified sections have been highlighted in yellow for easier identification).

To facilitate your review, we summarize the core updates of this rebuttal as follows:

1. Enhanced Rigorous Fair Comparisons and Statistical Significance (Addressing concerns on fairness and parameter counts)

Using PointAttN as an example, we supplemented strictly controlled experiments including equal-budget feedforward, generic recurrent, and baseline-equivalent pre-training setups. The results conclusively demonstrate that the performance improvement originates from BRICE's unique "dynamic predictive coding mechanism," rather than a mere expansion of parameter capacity or pre-training. For results with smaller margins of improvement, we have provided the mean and standard deviation (mean±std) across 3 different random seeds, proving the reliability of the gains through stable statistical variance. We have also detailed the rationale behind uniformly adopting the CRA-PCN training protocol to ensure absolute fairness when comparing baselines.

2. Clarified Methodological Originality and Comparisons with Related Work (Addressing questions regarding FBNet and Predify)

We refined the overly absolute statements in the original manuscript and provided a comprehensive, in-depth analysis—both in the main text and our responses—of the fundamental mechanistic differences between BRICE and FBNet regarding their operational domains (implicit feature space vs. physical coordinate space) and information loops. We added direct quantitative comparative experiments with FBNet and explicitly defined our relationship with Predify, highlighting BRICE's original contribution to cross-modal and cross-dimensional feature self-correction.

3. Rigorous Mathematical Definitions and Training Details (Addressing doubts regarding equations and theory)

We corrected the typographical error in Equation 7 caused by oversimplification in the original draft, explicitly clarifying that in actual implementation, BRICE fully retains the original multi-stage decoder losses of all baseline networks. We provided a rigorous mathematical definition for the feature Chamfer Distance and theoretically explained how the primary task loss LDecoderL_{Decoder} acts as a strong anchor to effectively prevent feature collapse (trivial solutions).

4. In-depth Discussion on the Module's Applicability Boundaries and Limitations

Incorporating your keen observations, we have thoroughly analyzed in the manuscript the root causes of performance degradation—specifically, feature misalignment—when BRICE is paired with decoders that possess extremely strong geometric hard priors (e.g., the symmetry prior in SymmCompletion) or dictionary guidance (e.g., ODGNet). We have candidly defined the optimal operating conditions for this plug-and-play module.

Furthermore, to provide a more intuitive understanding of our motivation for proposing BRICE, its design details, and the evolutionary process of its dynamic feature manifold, we have included an explanatory video in the supplementary materials. We kindly invite you to review it at your convenience.

We believe that through this detailed rebuttal and revision, the scientific rigor, transparency, and theoretical depth of the paper have been qualitatively elevated. We hope our sincere efforts and solid empirical data will fully address your concerns. We once again thank all reviewers for their immense effort in improving the quality of this manuscript, and we hope our work earns your approval.

Best regards,

Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.