{"id":"3b40bf8b-318c-457b-b9f9-337acff58dd4","arxiv_id":"2504.12643","paper_version":3,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"RoPETR claims 70.9% NDS on nuScenes by applying M-RoPE rotary embeddings to BEV coordinates and temporal IDs in StreamPETR, but a missing RayDN ablation leaves the attribution unclear.","lead":"RoPETR reports a new state-of-the-art NuScenes score for camera-only 3D detection by adding a rotary position embedding to StreamPETR. The paper does not isolate this embedding from a second module it also adds, so the improvement cannot be attributed.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The M-RoPE contribution is confounded by the concurrently added RayDN module: no StreamPETR+RayDN baseline or M-RoPE-only ablation is reported, so the NDS and mA VE gains cannot be attributed to the rotary embedding.","rationale":"The reader's weakest assumption is exactly the confound created by the one-sentence addition of RayDN in Section IV-B. In a paper whose headline contribution is a position embedding, the entire causal story depends on the controlled comparison that isolates that embedding. Because the model labeled RoPETR contains both the proposed M-RoPE and the RayDN module, the internal evidence does not distinguish between two hypotheses: (1) the rotary embedding improves velocity estimation, or (2) RayDN, a published module that Table II already shows improves NDS and mA VE over StreamPETR, accounts for the gap. The test-set comparison with RayDN is suggestive but inconclusive: the NDS gap is only 0.4, while mAP is lower, so the gap rests entirely on NDS submetrics that RayDN also targets. The missing ablation is therefore not a stylistic preference; it is the minimum experiment required to support the abstract's causal sentence. I do not see a reason to doubt the reported numbers as measurements; the problem is that they do not measure the proposed method in isolation. A controlled ablation could settle this. If the ablation were run and showed M-RoPE-only matching or exceeding the RoPETR results, the paper would still need equations and frequency details to be reproducible, but the central claim would be on much stronger ground. Given the missing evidence, the rejection recommended by the reader remains appropriate, and the verdict should stay REJECT.","tokens_in":7606,"tokens_out":8711,"duration_ms":85522,"concrete_test":"Train two additional configurations on the nuScenes val split under the Section IV-B settings, including the same 24-epoch, cosine-annealing, no-CBGS recipe on 8x A100 GPUs: (A) StreamPETR V2-99 320x800 with RayDN but without M-RoPE, and (B) StreamPETR V2-99 320x800 with M-RoPE but without RayDN. Compare NDS, mAP, and mA VE with the RoPETR val row (61.4 NDS, 52.9 mAP, 0.229 mA VE) and the StreamPETR baseline (57.1 NDS, 48.2 mAP, 0.263 mA VE). If configuration (A) is statistically indistinguishable from RoPETR, or if configuration (B) shows no mA VE improvement over StreamPETR, the reported gains are not attributable to the rotary embedding and the central claim fails. Ideally repeat the same pair on the test-set ViT-L setting using the Table II recipe.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's central claim is that the proposed M-RoPE-based positional embedding improves velocity estimation and pushes nuScenes NDS to state-of-the-art. Section IV-B states, \"Additionally, we incorporate the RayDN [16] module into our framework to further enhance detection performance.\" This means the reported RoPETR and RoPETR-e results are for StreamPETR plus two concurrent additions: the proposed embedding and RayDN. Table II shows RayDN (ViT-L) alone at 68.6 NDS and 0.220 mA VE, while RoPETR (ViT-L) is 69.0 NDS and 0.163 mA VE; RoPETR's mAP is actually lower than RayDN's (61.9 vs 63.1). No StreamPETR+RayDN baseline is reported on either the test or validation split, and no M-RoPE-only variant is reported. Therefore the difference between 68.6 and 69.0 NDS, and especially the mA VE drop from 0.220 to 0.163, could be caused entirely by the RayDN module, which already improves these numbers relative to StreamPETR (0.236 mA VE). The same confound applies to the validation table: the RoPETR row (61.4 NDS, 0.229 mA VE) is not compared against a StreamPETR+RayDN row, so the sentence \"clearly demonstrating the efficacy of our rotational positional encoding approach\" is not supported by the experiments. The report also omits the equations and frequency values that would define M-RoPE, but the attribution problem is the decisive gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RoPETR, a modification of the StreamPETR camera-only 3D detector. The claimed contribution is a Multimodal Rotary Position Embedding (M-RoPE) applied to query-key operations, intended to improve temporal modeling and, in particular, velocity estimation on the nuScenes benchmark. The authors also state that they incorporate the RayDN module into the same framework. Experiments report an NDS of 69.0% (mAP 61.9%) on the nuScenes test set with a ViT-L backbone, and 70.9% NDS (mAP 64.8%) with higher resolution and test-time augmentation, which the paper describes as state-of-the-art for camera-only 3D object detection. A validation-set comparison is also provided for a V2-99 backbone, where RoPETR reaches 61.4% NDS versus 57.1% for StreamPETR. The manuscript is short, method details are given only in prose, and no ablation study is included.","tokens_in":7900,"tokens_out":3399,"duration_ms":33687,"significance":"If the reported results are correctly attributed, the paper would represent a meaningful practical advance: improving velocity estimation is widely recognized as a bottleneck for NDS on nuScenes, and the reported mA VE reductions (0.236 to 0.163 in the test table) are substantial. The problem framing is sensible, and the use of rotary position embeddings for spatiotemporal queries is a plausible direction. However, the manuscript currently provides no formal specification of the proposed embedding, no ablations, and no baseline that separates the effect of M-RoPE from the concurrently added RayDN module. As a result, the central claim that M-RoPE drives the improvement is unsupported by the presented evidence. The paper's strengths are its clear motivation and the use of a standard benchmark, but these do not compensate for the missing isolation of the contribution.","major_comments":[{"comment":"The reported RoPETR results are confounded with the RayDN module. Section IV-B states, \"Additionally, we incorporate the RayDN [16] module into our framework,\" so every RoPETR row presumably includes both M-RoPE and RayDN. Table II shows RayDN (ViT-L) alone achieving 68.6 NDS and 0.220 mA VE, while RoPETR achieves 69.0 NDS and 0.163 mA VE. No StreamPETR+RayDN baseline and no M-RoPE-only (without RayDN) variant are reported. Since RayDN alone already improves NDS and mA VE over StreamPETR, the differences in Table II cannot be attributed to M-RoPE. The sentence in Section IV-C claiming that the results \"clearly demonstrat[e] the efficacy of our rotational positional encoding approach\" is therefore not supported by the experiments.","section":"Section IV-B, Table II"},{"comment":"The proposed method is not formally specified. No equations are provided for the frequency vectors omega and omega_t, the rotation angles theta_x, theta_y, and theta_t, or the exact mechanism by which these rotations are applied to Q/K in self-attention and cross-attention. The phrase \"multiplied by a logarithmic frequency vector\" leaves unclear whether the frequencies are fixed constants, learned parameters, or a combination. The reference to [26] (Qwen2-VL) does not identify a specific section or formula, and the adaptation to StreamPETR is described only qualitatively. This makes the contribution impossible to reproduce or to verify as a distinct technical novelty.","section":"Section III-B"},{"comment":"The validation evidence suffers from the same attribution problem. Table I compares RoPETR (V2-99) with StreamPETR and Stream3DPPE, but if the RoPETR row includes RayDN, as Section IV-B indicates, then the +4.3 NDS improvement over StreamPETR may be partly or wholly due to RayDN rather than M-RoPE. No validation row for RayDN and no ablation with and without M-RoPE are provided. Additionally, Table I uses a different backbone than the test-set SOTA claim, so the test results are not accompanied by a validation-table counterpart that would help assess generalization of the proposed component.","section":"Table I, Section IV-C"}],"minor_comments":[{"comment":"The abstract reports a state-of-the-art NDS of 70.86%, while Table II gives 70.9%; these values should be made consistent.","section":"Abstract and Table II"},{"comment":"The citation \"ViT-L [27] [20]\" for the ViT-L backbone is incorrect: [20] is SOLOFusion, not the ViT paper. The intended reference appears to be [4] (Dosovitskiy et al.).","section":"Section IV-C"},{"comment":"The name \"Multimodal Rotary Position Embedding\" is used, but the text does not explain what \"multimodal\" refers to in this application, given that only camera inputs are used. The relationship to the M-RoPE of Qwen2-VL should be clarified.","section":"Section III-B"},{"comment":"RoPETR (ViT-L) has a lower mAP than RayDN (61.9 vs 63.1) while having a higher NDS. The paper does not discuss this trade-off, which is relevant because the stated goal is to improve velocity estimation without harming detection accuracy.","section":"Table II"},{"comment":"The sentence \"Prior methods such as DETR3D and MV2D reported moderate results of 47.9% and 51.4% NDS\" lists numbers that correspond to specific input resolutions in Table II; the text should state the configurations to avoid misleading comparisons.","section":"Section IV-C"}],"recommendation":"reject","confidential_remarks":"The manuscript reads more like a workshop-style technical report than a complete journal submission. The decisive issue is the missing isolation of M-RoPE from RayDN; this is fixable in principle by adding a StreamPETR+RayDN baseline and a no-RayDN RoPETR ablation, but as submitted the central claim is not supported. The lack of formal definitions for the embedding further weakens the contribution. The authors may wish to resubmit after substantially extending the experiments and the method specification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the paper is a benchmark claim with a load-bearing confound. The authors take Qwen2-VL's M-RoPE, decompose it into temporal and 2D BEV components, plug it into StreamPETR's attention, and report a new nuScenes test NDS of 70.9%. The numbers are internally consistent and the validation table shows a healthy jump over the StreamPETR and 3DPPE baselines.\n\nCredit where due: the report is honest about borrowing M-RoPE from Qwen2-VL, and it identifies a real bottleneck (velocity estimation) that matters for planning. It also makes a sensible specific choice: normalizing BEV coordinates to [0,1] and encoding frame IDs separately.\n\nThe problem is that the method section contains no equations and no frequency values, and the experiments never isolate the proposed embedding from the concurrently added RayDN module. Section IV-B says \"Additionally, we incorporate the RayDN [16] module into our framework.\" The test table shows RayDN alone at 68.6 NDS and 0.220 mA VE; RoPETR is 69.0 NDS and 0.163 mA VE. Without a StreamPETR+RayDN baseline, the 0.4 NDS and 0.057 mA VE differences cannot be assigned to the rotary embedding. RayDN already improves mA VE from 0.236 to 0.220, so the velocity gain might be entirely RayDN's doing. The same confound applies to the validation table, and the sentence claiming \"clear demonstration of the efficacy of our rotational positional encoding approach\" is not supported by the data.\n\nThe paper also omits the actual definitions of the frequency vectors omega and omega_t, which are the only free parameters. That is a minor issue by comparison, but it makes the method hard to reproduce.\n\nThe reader's REJECT is the right call for the current version, but this is not a desk-reject case for the community. It is a short technical report with plausible SOTA numbers on a widely used benchmark. A competent referee could request the missing ablation and equations in one round. If the ablation shows the embedding helps beyond RayDN, this becomes a useful incremental contribution.","headline":"Benchmark claim with a load-bearing confound: no ablation isolates the proposed rotary embedding from RayDN.","tokens_in":8498,"tokens_out":3481,"would_cite":false,"duration_ms":28980,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"RoPETR claims that decomposing rotary position embeddings into spatial and temporal components improves temporal modeling enough to set a new camera-only 3D detection benchmark, reaching 70.9% NDS on the nuScenes test set.","keywords":["camera-only 3D object detection","rotary position embedding","temporal modeling","velocity estimation","nuScenes","StreamPETR","multi-view detection","autonomous driving"],"falsifier":"Train StreamPETR with RayDN but without the rotary embedding, using the same 24-epoch recipe; if its NDS and mAVE match RoPETR's 69.0 NDS and 0.163 mAVE, then the rotary embedding's claimed contribution is not supported.","tokens_in":1510,"feed_emoji":"🚗","tokens_out":7481,"duration_ms":116723,"temperature":0.7,"pith_summary":"This technical report targets velocity estimation as the main bottleneck in camera-only 3D object detection, the component that keeps the nuScenes Detection Score below mean average precision in query-based detectors. Building directly on StreamPETR, the paper adds a multimodal rotary position embedding that encodes each object's bird's-eye-view coordinates and its frame time as rotation angles applied inside attention. The claimed result is a jump to 61.4 NDS on the validation set and a new high of 70.9 NDS on the test set with an enhanced variant. The reason this matters is that a parameter-light change to position encoding, not a new depth network or dense representation, is said to close much of the velocity gap.","feed_headline":"Rotary position embedding lifts camera-only 3D detection to 70.9 NDS","feed_subtitle":"M-RoPE turns object location and frame time into rotation angles, cutting velocity error and raising nuScenes NDS.","key_machinery":"The central mechanism is Multimodal Rotary Position Embedding (M-RoPE), a rotary position embedding that turns normalized bird's-eye-view coordinates $(x,y)$ and a normalized frame time $t$ into rotation angles via log-frequency vectors, then applies channel-pair rotations to queries and keys in both self-attention and cross-attention. This gives the transformer an explicit relative-position and relative-time signal, which is what lets StreamPETR's propagated object queries encode motion dynamics rather than only appearance and depth.","core_discovery":"Multimodal Rotary Position Embedding (M-RoPE) normalizes each object's BEV center $(x,y)$ to $[0,1]^2$, multiplies it by a logarithmic frequency vector $\\omega$ to obtain rotation angles $\\theta_x$ and $\\theta_y$, and separately encodes a normalized frame identifier $t \\in [0,1]$ with another frequency vector $\\omega_t$ to obtain $\\theta_t$. These angles rotate channel pairs in query/key operations in self-attention and cross-attention, so attention becomes explicitly sensitive to relative BEV offsets and relative time between frames. In StreamPETR's streaming decoder, which already propagates object queries over time, the paper reports that this lowers mean velocity error from 0.220 to 0.163 on the test set and raises NDS from 68.6% to 69.0%, with the larger RoPETR-e variant reaching 70.9% NDS and 64.8% mAP.","pith_inferences":["The paper's attribution of the gain specifically to M-RoPE is not directly tested, because the RayDN module is added in the same step; a dedicated ablation would make the causal claim exact.","If relative-time encoding is the operative ingredient, varying the temporal frequency scale $\\omega_t$ should shift the velocity-error distribution in a predictable way, which is a direct behavioral test.","The same rotary decomposition is detector-agnostic and could be applied to other streaming query-based detectors, where the same velocity bottleneck likely appears."],"forward_implications":["If the central claim is right, the main measurable effect is a large drop in mean velocity error: RoPETR reports a test mAVE of 0.163 versus 0.220 for the RayDN baseline.","NDS can rise without mAP rising: RoPETR reports 69.0 NDS with 61.9 mAP against 68.6 NDS with 63.1 mAP for RayDN, so the gain is concentrated in motion-related metrics.","The embedding introduces no new depth network or dense supervision, so the reported gain does not come from added inference cost.","The method stacks with higher resolution and test-time augmentation, giving the enhanced RoPETR-e result of 70.9 NDS and 64.8 mAP on the test set."],"supporting_citations":[{"why":"StreamPETR is the base detector that RoPETR modifies, and its streaming object-query decoder is what the rotary embedding plugs into.","marker":"[27]"},{"why":"Supplies the M-RoPE rotary position embedding mechanism that the paper adapts to spatial and temporal coordinates.","marker":"[26]"},{"why":"The RayDN module is added into the framework; its lack of a separate baseline is load-bearing for attributing the gain.","marker":"[16]"},{"why":"3DPPE provides the point-aware positional embedding used as the main alternative comparison in the validation experiments.","marker":"[23]"},{"why":"The nuScenes dataset defines the NDS and mAP metrics and the benchmark on which all results are measured.","marker":"[1]"},{"why":"PETR establishes the 3D position embedding paradigm for multi-view detection that M-RoPE reformulates into a rotary form.","marker":"[17]"}],"fun_headline_variants":["RoPETR: M-RoPE cuts velocity error, hits 70.9 NDS","Rotary position embedding lifts camera-only 3D NDS to 70.9","Velocity error slashed, NDS 70.9 with M-RoPE in stream","M-RoPE: temporal rotary embedding for 70.9 NDS in 3D","RoPETR: from 68.6 to 70.9 NDS with rotary time encoding"],"cache_read_input_tokens":10496,"weakest_assumption_plain":"The paper's attribution of the NDS gain to the proposed embedding assumes that the concurrently added RayDN module does not account for the improvement.","fun_headline_variants_meta":{"raw":{"variants":["RoPETR: M-RoPE cuts velocity error, hits 70.9 NDS","Rotary position embedding lifts camera-only 3D NDS to 70.9","Velocity error slashed, NDS 70.9 with M-RoPE in stream","M-RoPE: temporal rotary embedding for 70.9 NDS in 3D","RoPETR: from 68.6 to 70.9 NDS with rotary time encoding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000229,"raw_usage":{"total_tokens":1440,"prompt_tokens":871,"completion_tokens":569,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":487,"completion_tokens_details":{"reasoning_tokens":450}},"tokens_in":487,"tokens_out":569,"duration_ms":5883,"temperature":1.0,"reasoning_tokens":450,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:26:24.880247+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train StreamPETR with RayDN but without the rotary embedding, using the same 24-epoch recipe; if its NDS and mAVE match RoPETR's 69.0 NDS and 0.163 mAVE, then the rotary embedding's claimed contribution is not supported.","supporting_citations":[{"cited_title":"Ray denoising: Depth-aware hard negative sampling for multi-view 3d object detection","cited_arxiv_id":null,"evidence_quote":"The RayDN module is added into the framework; its lack of a separate baseline is load-bearing for attributing the gain."},{"cited_title":"Petr: Position embedding transformation for multi-view 3d object detection","cited_arxiv_id":null,"evidence_quote":"PETR establishes the 3D position embedding paradigm for multi-view detection that M-RoPE reformulates into a rotary form."}],"review_version":1}