acceptodds
Withdrawn from ICLR 2027

TGVPose: Lightweight Category-Level Object Pose Estimation via Topology Graph-Guided Adaptive Graph Vector Networks

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Object pose estimation serves as a crucial approach to enabling reliable robotic grasping. To achieve high generalization in real-world scenarios, existing pose estimation methods commonly adopt category-level object pose estimation strategies. However, current high-precision methods often suffer from large parameter volumes, rendering them difficult to deploy on computation-constrained edge devices such as mobile robots and unmanned aerial vehicles (UAVs). To address this issue, this paper proposes TGVPose, a high-precision and lightweight category-level object pose estimation method via topology graph-guided adaptive graph vector networks. First, to overcome the fixed-graph limitation inherent in conventional graph vector networks, we introduce an Adaptive Multi-Relation Gated Vector Neuron Layer (AMG-VN), which enables the network to adaptively extract both local and global features. Second, to enhance the network's understanding of object shapes, we present a category-prototype persistent topology conditioning method, thereby achieving accurate object pose estimation. Finally, we evaluate TGVPose across multiple datasets and real-world object pose estimation tasks. The experimental results demonstrate that TGVPose effectively balances accuracy and lightweight efficiency, delivering superior pose estimation performance compared to existing lightweight methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Category-level object pose estimation

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_SR4rOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes TGVPose, a lightweight object pose estimation method. It uses adaptive graph vector networks and topology conditioning. Experimental results indicate that this method can achieve faster inferences with comparable detection accuracy.

Critical strengths
  • The proposed method looks reasonable and the inference speed is improved a lot, especially on the edge device.
  • This paper provided ablation studies to demonstrate the effect of each module.
Critical weaknesses
  • The FPS was increased with sacrificing the pose estimation accuracy. Compared to SEGPose, mAP of TGVPose for IoU75IoU_75 and 10°5cm10^{\degree}5cm regressed 10+%. I have a concern about such a large regression. Moreover, SEGPose can already achieve 13.4 FPS, I feel this FPS is high enough for robotic grasping task since objects are static.
  • Although the paper provides ablation studies about mAP, there is no qualitative visualizations about the proposed modules. For instance, without visualization, it is a little difficult to understand the real impact of topology-guided point cloud reconstruction and topology conditioning.
Questions for authors
  • Could the authors explain why higher FPS is required for robotic grasping task? Especially when objects are static? How do we balance the accuracy-speed trade-off?
  • Could the authors provide additional visualizations of the proposed modules, such as topology-guided point cloud reconstruction and topology conditioning?
Reviewer_GnHeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

To resolve the problem that current high-precision methods often suffer from large parameter volumes, rendering them difficult to deploy on computation-constrained edge devices such as mobile robots, this paper proposes TGVPose, a high-precision and lightweight category-level object pose estimation method via topology graph-guided adaptive graph vector networks. It contains two major modules, i.e. an adaptive multi-relation Gated Vector Neuron Layer (AMG-VN), a category-prototype persistent topology conditioning module. Its experimental results demonstrate that TGVPose effectively balances accuracy and lightweight efficiency, comparing existing lightweight methods.

Critical strengths
  1. It is an innovative idea by combining category-prototype persistent topology with 6D pose estimation.
  2. This paper also did many experiments on both two public real datasets and many latest RGBD-based and depth-only pose estimation.
Critical weaknesses
  1. In the table 1, the author should state clearly: 1) whether the accuracy results are copied from their original paper or re-run by the author; 2) especially for the FPS data, if copied, what are the hardware settings of each method, since it is a fair comparison only all methods run on the same hardware configuration; 3) the benefits of the proposed method is not so significant, e.g. the TG-Pose method which have much higher accuracy while his FPS is also 50.0, which is comparable to your 73.7 FPS; 4) the analysis for proposed method is worse than RGB-D SOTA is “TGVPose relies solely on depth information, without utilizing RGB images to assist the network in extracting richer object details.”, I think such kind of reason cannot support very well, since your method can also using RGB data.
  2. Similar case, i.e. the benefits of the proposed method is not so significant, appears in Table 2. For example, the accuracy of SEGPose method is better than the newly proposed method.
  3. The Figure 1 is not referenced in the paper. In addition, why the method TG-Pose is not listed on the figure? Also, what are the prior-based and prior-free methods since those pose estimation methods are divided into RGBD-based and depth-only methods in table 1 and table2.
Questions for authors
  1. the manuscript should be re-organized in a good-reading format. For example, in the method over section, it is hard to get the method over, how many modules it contains and what their relations?
  2. The related work section is too general, in my opinion, it should focus more on lightweight pose estimation methods, to emphasize your major contribution.
Reviewer_waBkOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

TGVPose is a lightweight, category-level object pose estimation network that relies solely on depth point clouds. It employs an adaptive multi-relation gated vector neuron layer to fuse local, invariant, and global features, while enhancing the invariant features with category topology priors computed offline. Although its accuracy is slightly lower than that of SEGPose on the REAL275 dataset, it is 1.6–2.3 times faster and achieves the highest success rate in real-world grasping tasks on the UR3 robot.

Critical strengths

The efficiency evaluation is grounded in real-world edge devices. The paper actually deploys the model on the Jetson platform to measure inference speed and benchmarks it against the prior work on the same hardware. This provides a direct answer to the question of "whether it can run in real-time on mobile robots" offering a more meaningful benchmark for future research.

Critical weaknesses

1.The core claim of this paper is to trade acceptable accuracy loss for real-time inference on edge devices, but this claim lacks sufficient support. The paper does not explain where the acceleration actually comes from: the proposed layer is structurally more complex than the single-graph vector neuron layer, yet the overall model is faster. The causal link is missing. Furthermore, "lightweight" is presented solely in terms of frame rate, without metrics such as parameter count, FLOPs, or memory footprint to demonstrate whether the model itself is actually lighter. The deployment conditions are also not specified, making the efficiency conclusions difficult to attribute and reproduce. More importantly, the paper provides only a single operating point without showing the accuracy-latency trade-off curve. Therefore, it is impossible to determine whether this point lies on the Pareto frontier or if it outperforms a simply scaled-down version of the prior work.

2.There is no comparison with traditional compression methods. The paper treats architectural design as the sole path to lightweighting, yet it completely fails to compare against traditional compression techniques such as quantization, pruning, distillation, or inference engine optimization. Nor does it specify its own deployment settings. This omission directly undermines the paper's motivation: the starting point is that the prior work cannot run in real-time on edge devices. However, applying standard compression and deployment optimizations to the prior work would typically yield significant acceleration with negligible accuracy loss. If that is sufficient to achieve real-time performance, the accuracy sacrificed for speed in this work becomes unnecessary. The authors might argue that architectural improvements are orthogonal to compression and can be stacked, which is theoretically sound. However, orthogonality must be demonstrated with data from both before and after compression on the same device, rather than by mere assertion. In the absence of such comparisons and model scale metrics, readers cannot judge whether the proposed approach offers a genuine advantage over zero-research-cost engineering methods.

3.The contributions are incremental and minor. The paper provides no analysis beyond ablation studies to explain why the components are effective. Every design choice in the method lacks justification at a principled level, and there is no comparison with alternative approaches of equivalent cost. The argument for equivariance merely checks existing designs rather than serving as a basis for deriving the new design. The ablation section offers mostly narrative, speculative mechanistic explanations for the results, without any direct evidence to support them. Moreover, the ablation study itself lacks multiple runs and variance reporting; many of the observed gaps fall within the range of reasonable noise, which casts doubt on the credibility of the "verification." In terms of novelty, each component is a recombination of existing ideas. The topology prior was already used in the authors' own previous work; the modification here is mainly to make it offline and pair it with empirical constants. Thus, the methodological contribution is an incremental engineering tweak rather than a new design principle.

Questions for authors

see weakness.

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

This paper proposes TGVPose, a lightweight framework for category-level object pose estimation. The overall pipeline takes RGB-D observations as input, using Mask R-CNN for instance segmentation and category recognition, followed by depth-based point cloud extraction for pose estimation. The method aims to balance estimation accuracy and computational efficiency for resource-constrained robotic platforms. The main contributions include an Adaptive Multi-Relation Graph Vector-Neuron (AMG-VN) layer, which integrates local geometric neighborhoods, rotation-invariant feature similarities, and global context through adaptive gating while preserving rotation equivariance. Additionally, a persistent topology conditioning module incorporates category-level shape priors derived from persistent homology into the learned features without requiring online topological computation. The authors evaluate TGVPose on the REAL275 and Wild6D benchmarks, conduct ablation studies, and report inference efficiency on an RTX 3090 GPU and an NVIDIA Jetson platform. Real-world grasping experiments are also presented using a UR3 robot. The reported results suggest an accuracy-efficiency trade-off compared with existing methods.

Critical strengths
  1. The paper addresses a practically relevant problem: efficient category-level object pose estimation for resource-constrained robotic platforms. The reported results show substantially improved inference speed compared with SEGPose, particularly on NVIDIA Jetson, while maintaining reasonably competitive pose estimation accuracy. This suggests potential value for deployment in robotic applications.
  2. The proposed AMG-VN module combines multiple neighborhood relations through rotation-invariant adaptive gating while preserving equivariance. The topology conditioning module further introduces category-level structural priors without requiring online persistent-homology computation. Although the individual techniques are not fundamentally new, their integration provides a plausible approach to balancing geometric representation quality and computational efficiency.
Critical weaknesses
  1. Supplementary Table 4 explicitly states: "These reference values must be replaced by measurements obtained using the same split, training schedule, and hardware." This strongly suggests that the reported values are illustrative placeholders rather than verified experimental measurements. Importantly, this table contains critical comparisons against learned category embeddings, fixed topology prototypes, and category-only predictors, which are essential for establishing whether the proposed topology conditioning provides benefits beyond ordinary category information. Without reliable experimental evidence for these comparisons, a central methodological claim remains insufficiently supported.
  2. The real-robot grasping results contain substantial numerical contradictions. Supplementary Table 7 reports success rates of 80.0%, 71.1%, 73.3%, and 57.8% across four tasks, corresponding to 127 successful trials out of 180. However, the accompanying textual descriptions report 88.9%, 82.2%, 75.6%, and 64.4%, corresponding to 140 successful trials out of 180. These results cannot simultaneously describe the same experiments. Additionally, the reported 5°/5cm performance of TGVPose varies between 55.2% and 56.1% across the main paper and supplementary material, without a sufficiently clear explanation of the corresponding model configurations. These inconsistencies go beyond minor presentation issues and raise substantial concerns about the reliability, reproducibility, and validity of the experimental evidence supporting the paper's conclusions.
Questions for authors

Q1. Supplementary Table 4 explicitly states that its reference values "must be replaced by measurements." Are these values illustrative placeholders or actual experimental results? Could the authors clarify their origin and provide the corresponding measured results? Q2. Supplementary Table 7 reports 127 successful trials out of 180, whereas the accompanying descriptions in Section 6 correspond to 140 successful trials. Which results are correct? Could the authors provide the actual success counts for each task and explain the discrepancies? Q3. The reported 5°/5cm performance differs between 55.2% and 56.1%. Furthermore, Supplementary Section 4 introduces a category-free topology predictor with instance-level supervision that differs from the main method. Could the authors clarify which configuration corresponds to each result, which is the final TGVPose model, and whether the accuracy and efficiency comparisons use the same configuration? Q4. The paper accounts for rotational symmetry in the training objectives by masking the secondary rotation-axis supervision for symmetric objects. However, neither the main paper nor the supplementary material clearly specifies how rotational symmetry is handled during evaluation. Could the authors clarify whether the reported rotation-based mAP metrics (e.g., 5°/2cm and 5°/5cm) and 3D IoU metrics follow a symmetry-aware evaluation protocol, such as the official NOCS evaluation protocol? In particular, how are symmetry-equivalent rotations handled for continuously rotationally symmetric objects?

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

Related papers

Discussion (0)

Sign in to comment.