{"id":"fe857b84-7f2d-43d4-a754-a449155d2f1e","arxiv_id":"2412.18775","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"ObitoNet fuses ViT image tokens and FPS/KNN point-cloud tokens with a cross-attention decoder, reporting a Chamfer loss of 1.36 on Tanks and Temples versus PointMAE's 1.53.","lead":"ObitoNet combines image and point cloud data using attention to reconstruct 3D shapes. The authors report a lower Chamfer loss than PointMAE, but the evidence is thin and the comparison is unconvincing.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unvalidated CLIP-based image–point-cloud pairing (§3.2) is the load-bearing assumption; without the threshold and pose-alignment check, Table 4 cannot support the multimodal fusion claim.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the CLIP-based pairing in §3.2 is unspecified and unvalidated. My independent read of the full text confirms this is the most consequential soft spot. All other issues—missing error bars, absent dataset splits, the duplicate arXiv reference, artifact-laden writing—are serious but secondary: even a perfectly reported experiment would rest on the pairing step. If the pairs are misaligned, the cross-attention mechanism cannot learn meaningful image-to-geometry relationships, so the central claim collapses regardless of how Table 4 is formatted. A concrete alignment check can settle the question: the released code either contains the threshold and produces pose-consistent pairs, or it does not. I do not see a way to rescue the central claim without that evidence. The reader's verdict of REJECT is therefore appropriate, and my stress-test does not move it.","tokens_in":6546,"tokens_out":2306,"duration_ms":23933,"concrete_test":"Run the released GitHub dataset-generation code on Tanks and Temples, extract the exact CLIP threshold and projection parameters, then sample 100 generated pairs. For each pair, register the matched RGB image against the projected point-cloud reference (e.g., with SfM poses from the dataset or manual keypoint correspondences) and compute median reprojection error. If the threshold is absent from the code, or if median reprojection error exceeds a few pixels for a substantial fraction of pairs, the pairing assumption fails and Table 4 must be recomputed on verified pairs.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that ObitoNet's cross-attention fusion of image and point-cloud tokens achieves Chamfer 1.36 on Tanks and Temples, beating PointMAE's 1.53. For this claim to be true, the paired image and point cloud must actually represent the same scene geometry from a compatible viewpoint, because the cross-attention block uses point-cloud tokens as queries and image tokens as keys/values (§3.3). The dataset-construction step in §3.2 is the only source of such pairs, and it is not validated. The paper says images and point clouds are paired when a CLIP similarity score exceeds a 'predefined threshold' but never reports that threshold, never describes how the point cloud is projected onto a 2D plane, and never measures whether the matched RGB image and the projected point cloud agree geometrically. Tanks and Temples images are captured from distinct camera poses; a scene-level CLIP match need not imply POV alignment. If many pairs are misaligned, the cross-attention module trains on spurious image–geometry correspondences, and the reported Chamfer values reflect dataset artifacts rather than genuine multimodal reconstruction. The reference-list inconsistency ([5] and [6] share arXiv:2111.14819) and the missing error bars in Table 4 are additional reporting failures, but the pairing assumption is the one whose failure would invalidate the method's core mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ObitoNet, a multimodal point cloud reconstruction framework that fuses Vision Transformer (ViT) image tokens with point cloud tokens derived from Farthest Point Sampling (FPS) and K-Nearest Neighbors (KNN) through a cross-attention module. The authors describe dataset construction on Tanks and Temples, where CLIP is used to pair images with point clouds, and report Chamfer distances for three ObitoNet variants and a PointMAE baseline (Table 4), claiming comparable or better performance. The paper also includes model configuration tables and qualitative visualizations.","tokens_in":6728,"tokens_out":3460,"duration_ms":31720,"significance":"If the central empirical claim were properly supported, the cross-attention fusion of image and point cloud tokens for point cloud completion and upsampling would be a modest but useful contribution to multimodal 3D reconstruction. The paper gives credit for a modular architecture, explicit model configuration tables, a GitHub repository link, and an honest acknowledgment of the POV-alignment difficulty in Section 3.2. However, the evaluation as presented cannot sustain the claim of comparable performance: the CLIP-based pairing is unvalidated, the baseline comparison is minimal, and the experiments lack error bars, dataset splits, and the ShapeNet evaluation promised in the abstract. These are load-bearing weaknesses, not presentation issues.","major_comments":[{"comment":"The CLIP-based image–point-cloud pairing is load-bearing for the multimodal fusion claim, but it is never validated. Section 3.2 says images and point clouds are paired when a similarity score exceeds 'a predefined threshold' without reporting the threshold, describing the projection method beyond 'projected onto a 2D plane', or checking geometric/pose consistency. Since Section 3.3 uses point cloud tokens as queries and image tokens as keys and values, any misaligned pairs would train the cross-attention module on spurious correspondences. Until the threshold and a validation of POV alignment are provided, the Chamfer values in Table 4 cannot be interpreted as evidence for genuine multimodal reconstruction.","section":"§3.2 / Table 4"},{"comment":"The central claim of 'comparable performance to state-of-the-art' is unsupported by the experiments. Table 4 reports single Chamfer loss values for ObitoNet/Base, ObitoNet/Large, ObitoNet/ViTMAE, and PointMAE, but gives no error bars, no train/validation/test split, no number of test samples, and no source for the PointMAE number. The abstract claims evaluation on ShapeNet, but Section 4 uses only Tanks and Temples. With a single unquantified baseline, the reported margin (1.36 vs. 1.53) could be noise or a dataset artifact, so the comparison does not meet the standard needed to support the paper's headline claim.","section":"§4 / Table 4"},{"comment":"Experimental reproducibility is not met because key training details are missing. The paper gives some architectural parameters (token size, number of tokens, MAE decoder depth) but does not report learning rate, batch size, optimizer, masking ratio, KNN group size M, training epochs for the Base and ViTMAE variants, or how the test set is constructed. Without these details, the reported Chamfer losses are not reproducible, which further weakens the empirical contribution.","section":"§3.5 / §4.2"}],"minor_comments":[{"comment":"References [5] and [6] are both assigned arXiv:2111.14819, but [5] is VoxelGridNet and [6] is PointBERT; at least one identifier is incorrect.","section":"References"},{"comment":"The abstract states that experimental evaluations were performed on ShapeNet, but Section 4 uses only the Tanks and Temples dataset; the paper should clarify which dataset was actually used.","section":"Abstract vs. §4"},{"comment":"Equation (1) defines a squared L2 Chamfer loss, while the abstract mentions 'Chamfer Distance (L1/L2)'; please clarify whether the loss uses L2 squared, L1, or both.","section":"Equation (1)"},{"comment":"Figure 1 is captioned 'PointMAE pointcloud reconstruction sample outputs' but is not referenced in the text and its relationship to the proposed method is not explained.","section":"Figure 1"},{"comment":"The caption of Table 2 reads 'The results of this experiment are highlighted in Table 2,' but Table 2 lists model configuration parameters; the results appear in Table 4.","section":"Table 2 caption"}],"recommendation":"reject","confidential_remarks":"This manuscript appears to be an early draft that does not yet meet the evidentiary bar for publication. The unvalidated CLIP pairing (§3.2) and the single-number, single-baseline evaluation (Table 4) are fundamental problems: they directly undermine the paper's central empirical claim, and patching them would require substantial additional data, experiments, and validation. I therefore recommend rejection rather than major revision. I also note the internal inconsistency between the abstract's mention of ShapeNet and the actual Tanks and Temples experiments, and the duplicated reference identifiers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a straightforward combination of known components: ViT for images, FPS/KNN tokenization for point clouds, and cross-attention for fusion. That's not new, but it's a reasonable architecture, and the three-stage training order is sensible. The CLIP-based idea for pairing images and point clouds is genuinely interesting, because it addresses a real problem: Tanks and Temples doesn't come with aligned POV pairs. That part deserves credit.\n\nThe problem is that the pairing is the load-bearing assumption and it is never validated. Section 3.2 says pairs are matched when a CLIP similarity score exceeds a 'predefined threshold,' but the threshold isn't reported, and there is no check that a matched image and point cloud actually show the same geometry from a compatible viewpoint. The paper itself notes that KITTI is unsuitable because images cover only parts of the point cloud, so the authors know the issue. Yet they don't show that Tanks and Temples avoids it. Without that, the cross-attention module may be learning spurious image-to-geometry correspondences, and the numbers in Table 4 would reflect dataset artifacts rather than real multimodal fusion.\n\nTable 4 is also thin: single Chamfer values with no error bars, no dataset splits, no number of test scenes, and only one baseline, PointMAE. The abstract says ShapeNet; the experiments use Tanks and Temples. That discrepancy isn't fatal by itself, but it's sloppy. The shared arXiv ID between references [5] and [6] under different titles and authors is a careless error and raises questions about the citation handling.\n\nOn the positive side, the methodology is mostly coherent, and the paper doesn't overclaim wildly in the conclusion—it says 'comparable to PointMAE,' which is modest. But the evidence doesn't even support that modest claim. This is an incomplete student-project write-up, not a paper ready for peer review. I'd desk-reject it now, but the CLIP pairing idea is worth pursuing with proper validation in a future submission.","headline":"The core idea is plausible but nothing in the experiments supports the headline claim; the CLIP-based pairing is unvalidated and Table 4 is the only evidence.","tokens_in":7314,"tokens_out":2110,"would_cite":false,"duration_ms":20206,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Cross-attention fusion of image tokens with point-cloud tokens reconstructs point clouds at Chamfer 1.36, undercutting PointMAE's 1.53.","keywords":["point cloud reconstruction","multimodal fusion","cross-attention","Vision Transformer","farthest point sampling","K-nearest neighbors","Chamfer distance","Tanks and Temples"],"falsifier":"Use the Tanks and Temples camera poses to project each training point cloud into its paired RGB image and measure the median reprojection error; if a large share of the CLIP-matched pairs are misaligned, or if retraining with randomly shuffled image-point pairs reproduces the same Chamfer losses, the reported gain is not coming from genuine cross-modal fusion.","tokens_in":6264,"feed_emoji":"🧊","tokens_out":11969,"duration_ms":91081,"temperature":0.7,"pith_summary":"ObitoNet claims that a transformer-based point cloud reconstructor performs better when it can also see an RGB image of the same scene, with point cloud tokens attending to image tokens through cross-attention. On paired scenes from the Tanks and Temples dataset, the paper reports Chamfer distances of 1.36 for the base model and 1.20 for the large model, compared with 1.53 for the point-only PointMAE baseline. If the claim is right, it means sparse or noisy 3D scans can be densified and completed using ordinary images as guidance, which are cheaper to acquire than dense scans. The paper further claims that a three-stage training schedule, point-only completion followed by image-token adaptation and then joint fine-tuning, keeps each component reusable for downstream tasks.","feed_headline":"Image-plus-point fusion reaches Chamfer 1.36 beating PointMAE 1.53","feed_subtitle":"Each 3D token queries the image for cues that fill sparse geometry.","key_machinery":"The load-bearing object is the cross-attention module: point cloud tokens become the queries $Q$, image tokens become the keys $K$ and values $V$, so each geometric token retrieves semantic context from the image to fill missing geometry. The point cloud tokens are formed by FPS anchor selection and KNN grouping into clusters $N \\in \\mathbb{R}^{B \\times G \\times M \\times 3}$, then projected to token dimension; the image tokens are ViT patch embeddings with the CLS token removed. The fused token stream is refined by a transformer decoder with residual connections, and the reconstruction head maps features to 3D coordinates with two 1D convolutional layers and ReLU. The three-stage training order is part of the machinery: it first teaches point-only gap filling, then adapts the image encoder to the frozen point-cloud branch, and finally fine-tunes all modules jointly.","core_discovery":"On its own terms, the paper establishes that geometry completion can be driven by a cross-attention block in which point cloud tokens act as queries and image tokens act as keys and values. The point cloud tokenizer selects anchor centers with farthest point sampling and groups local neighborhoods with k-nearest neighbors, while the image branch embeds non-overlapping patches with a pretrained Vision Transformer and discards the CLS token. Fused tokens pass through a transformer decoder and a two-layer 1D convolution head that outputs three times as many points as were input. Trained in three stages on CLIP-matched image-point-cloud pairs, the base model reaches Chamfer distance 1.36, the large variant 1.20, and the ViTMAE variant 1.41, against 1.53 for the PointMAE baseline. The paper takes these numbers to show that multimodal fusion improves high-resolution reconstruction over point-cloud-only processing.","pith_inferences":["The Table 4 comparison is on CLIP-matched Tanks and Temples pairs, not on the ShapeNet benchmark named in the abstract; reproducing the method on ShapeNet or another standard split would test how far the numbers generalize.","A direct ablation that feeds random image-point pairs through the same cross-attention architecture would isolate whether the gain comes from genuine geometric fusion or from the point-cloud transformer alone.","The CLIP pairing threshold is left unspecified; publishing it and the pairwise similarity distribution would let others judge whether the training set is truly viewpoint-aligned.","The same cross-attention design is a natural fit for image-guided completion of partial LiDAR or depth scans in robotics, provided a calibration-based pairing replaces the similarity matching."],"forward_implications":["If the reported Chamfer numbers transfer beyond the tested pairs, image-guided cross-attention gives lower reconstruction error than the point-only PointMAE pipeline.","The 3x upsampling objective means the same fusion can densify sparse scans, not just reconstruct at the input density.","The modular tokenizers and cross-attention block can be reused for tasks the paper lists, including object classification, semantic segmentation, point cloud registration, and multimodal spatial reasoning.","The three-stage schedule allows components to be frozen and reused independently, lowering the barrier to adapting the model to new tasks."],"supporting_citations":[{"why":"supplies the pretrained Vision Transformer that produces the image tokens used in cross-attention.","marker":"[1]"},{"why":"supplies the self-attention formulation behind the cross-attention module and transformer decoder.","marker":"[2]"},{"why":"serves as the point-cloud-only baseline that ObitoNet replicates in training stage one and beats in Table 4.","marker":"[3]"},{"why":"provides the foundation for processing unordered point sets that the FPS/KNN tokenizer builds on.","marker":"[4]"},{"why":"shows masked point modeling with transformers, the pretraining context for the point-cloud branch.","marker":"[6]"},{"why":"introduces hierarchical FPS and KNN grouping that the point cloud tokenizer uses to form local clusters.","marker":"[10]"}],"fun_headline_variants":["ObitoNet fuses images and points for sharper 3D reconstruction","Cross-attention fills sparse point clouds with image cues","Image-guided point cloud reconstruction hits Chamfer 1.36","Point clouds query images to beat PointMAE in 3D","Multimodal fusion lifts 3D reconstruction from sparse data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each CLIP-matched image really shows the same scene from the same viewpoint as its paired point cloud; the paper calls the matching threshold predefined but never states it and reports no check of the pairs.","fun_headline_variants_meta":{"raw":{"variants":["ObitoNet fuses images and points for sharper 3D reconstruction","Cross-attention fills sparse point clouds with image cues","Image-guided point cloud reconstruction hits Chamfer 1.36","Point clouds query images to beat PointMAE in 3D","Multimodal fusion lifts 3D reconstruction from sparse data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000557,"raw_usage":{"total_tokens":2581,"prompt_tokens":809,"completion_tokens":1772,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":425,"completion_tokens_details":{"reasoning_tokens":1685}},"tokens_in":425,"tokens_out":1772,"duration_ms":11932,"temperature":1.0,"reasoning_tokens":1685,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:29:00.921217+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use the Tanks and Temples camera poses to project each training point cloud into its paired RGB image and measure the median reprojection error; if a large share of the CLIP-matched pairs are misaligned, or if retraining with randomly shuffled image-point pairs reproduces the same Chamfer losses, the reported gain is not coming from genuine cross-modal fusion.","supporting_citations":[],"review_version":1}