{"id":"d141704a-fc21-4697-b563-9e722f923223","arxiv_id":"2611.16793","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"EG360 is a 6.593M-parameter panorama depth estimation model that achieves real-time inference on embedded devices and competitive zero-shot accuracy via adaptive distillation and spherical feature fusion.","lead":"EG360 is a lightweight neural network that estimates depth from 360-degree panoramic images while running at 35 frames per second on an embedded Jetson Orin NX device. It achieves this by combining knowledge from two larger teacher models (one panoramic, one perspective) using an adaptive weighting scheme and a spherical attention module.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"No significant objection identified","rationale":"The reader's verdict is CONDITIONAL with high confidence, citing the lack of error bars, tau sensitivity analysis, and code availability. I agree that these are reasonable concerns for a reproducibility audit, but I do not find them load-bearing for the central claim. The paper's own ablation (Table 7) shows that simple averaging of teacher losses (which corresponds to tau → ∞) already achieves AbsRel 0.168 on Matterport3D, very close to the full APPD result of 0.169. This indicates that the exact value of tau is not critical to the method's success. The absence of error bars is a reporting weakness but does not make the central claim implausible. The code release is promised and would enable verification. The paper's limitations section honestly acknowledges the limited unlabeled data and temporal incoherence. I see no hidden assumption or internal contradiction. Therefore I recommend UNCHANGED: the CONDITIONAL verdict is appropriate, and my analysis does not suggest a need to strengthen or weaken it.","tokens_in":17792,"tokens_out":509,"duration_ms":8794,"concrete_test":"Run the APPD weighting with tau=0.1 and tau=1.0 on the Matterport3D validation set; if the AbsRel varies by more than 0.01 across this range, the sensitivity is higher than the paper implies, but even then the central claim of competitive accuracy at 35 FPS would not be overturned.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper presents a well-engineered system with thorough ablation studies, clear motivation for each design choice, and honest discussion of limitations. The central claim—that EG360 achieves competitive zero-shot panorama depth accuracy at 35 FPS on an embedded device—is supported by the reported numbers and the ablation tables. The reader's weakest_assumption about tau sensitivity is acknowledged but not a load-bearing concern: the paper states tau is set experimentally (Section 3.3), and the ablation in Table 7 shows that even simple averaging (which corresponds to tau → ∞) already yields strong results, with APPD providing a modest additional improvement. The absence of error bars is a reproducibility concern but does not invalidate the central claim. The code release promise addresses the main verification gap. I find no internal inconsistency or hidden assumption that would undermine the core argument.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper presents EG360, a lightweight monocular panorama depth estimation model designed for efficient zero-shot generalization on embedded platforms. The method combines an Adaptive Panorama-Perspective Distillation (APPD) strategy that dynamically weights supervision from a panorama teacher (DAP) and a perspective teacher (MoGe2) based on per-face planar residuals, and a Spherical Cross-Level Feature Fusion (SCFF) module that enhances global context via cross-attention with a spherical direction bias. The model is trained on labeled synthetic data and unlabeled real data with cached teacher pseudo-labels. Experiments on Matterport3D, Stanford2D3D, and Metropolis show that EG360 (6.593M params, 23.492 GFLOPs) achieves competitive zero-shot AbsRel (0.173, 0.094, 0.237) and runs at 35 FPS on a Jetson Orin NX, outperforming prior lightweight models and approaching larger foundation models in accuracy while being much faster.","tokens_in":18046,"tokens_out":1814,"duration_ms":14589,"significance":"If the reported results hold, EG360 represents a meaningful step toward practical deployment of panorama depth estimation on resource-constrained devices. The paper demonstrates that a carefully designed distillation strategy combining complementary teachers and a lightweight architecture with spherical-aware global context can narrow the gap between efficient models and large foundation models. The thorough ablation studies (Tables 4-10) provide evidence for the contribution of each component. The code release promise supports reproducibility. The work is likely to be of interest to the community working on 360° perception, efficient deep learning, and embedded vision systems.","major_comments":[{"comment":"The hyperparameter τ in the teacher weighting formula w_P = τ/(r_c+τ), w_C = r_c/(r_c+τ) is set to 0.3 experimentally (line after Eq. (3)). The paper does not report any sensitivity analysis for τ. Table 7 shows that simple averaging (which corresponds to τ → ∞) already yields strong results, and APPD provides a modest improvement on Stanford2D3D (AbsRel 0.101 vs 0.109) but slightly worse on Matterport3D (0.169 vs 0.168) and Metropolis (0.244 vs 0.238). Without a sensitivity study, it is unclear whether the chosen τ is robust across datasets or whether the improvement over averaging is statistically significant. The authors should either provide an ablation over τ values (e.g., τ ∈ {0.1, 0.3, 0.5, 1.0}) or justify why τ=0.3 is a principled choice.","section":"§3.3, Eq. (3)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The paper states that training runs for 100,000 iterations with a batch size of 8 for labeled data and 8 for unlabeled data. However, the ablation in Table 6 varies the unlabeled batch size while keeping the labeled batch size fixed. The paper does not specify whether the total number of iterations or the total number of epochs is held constant when varying the unlabeled batch size. If the number of iterations is fixed, then a smaller unlabeled batch size means fewer total unlabeled samples seen during training, which could explain the performance drop. The authors should clarify the training protocol and ensure that the ablation is fair.","section":"§4.1, Implementation Details","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The 'Manual' strategy (assign DAP to ZInD and MoGe2 to HoliCity) achieves AbsRel 0.166 on Matterport3D, which is better than APPD's 0.169. The paper does not discuss this result. If a simple dataset-level assignment outperforms the geometry-adaptive weighting on one dataset, the claim that APPD is 'guided by scene geometry' and provides a 'useful accuracy trade-off' needs qualification. The authors should explain why APPD underperforms Manual on Matterport3D and whether this indicates that the planar residual heuristic is not always optimal.","section":"§4.3, Table 7","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The SCFF ablation is performed on Stanford2D3D using labeled data only (no pseudo-labels). However, the full EG360 model uses both APPD and SCFF together. The interaction between the two components is not studied: does SCFF provide additional benefit when APPD is already applied? Table 4 shows that adding SCFF to the baseline improves AbsRel from 0.126 to 0.113 on Stanford2D3D, but when APPD is already present, adding SCFF (EG360 vs +APPD) improves from 0.101 to 0.094. The paper should discuss whether the gains from SCFF are additive or whether there is diminishing returns.","section":"§4.3, Table 10","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The comparison with foundation models (DAP, DepthMaster, PaGeR) is informative but the paper does not discuss the fact that these models are evaluated using their released weights, which may have been trained on different data distributions. The lightweight models (UniFuse, BiFuse++, ACDNet) are retrained on the same labeled data for fairness, but the foundation models are not. This asymmetry should be acknowledged as a limitation when interpreting the accuracy gap. Additionally, DepthMaster and PaGeR ran out of memory (OOM) on the Jetson, so their latency is not reported; this is fine, but the paper should note that the comparison is incomplete.","section":"§4.2, Table 3","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The abstract states 'our code will also be made public for future research.' It would be helpful to specify the intended license and repository platform.","section":"Abstract","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The normalized plane-fitting residual r_c is defined as the RMS distance to the fitted plane divided by the mean radial depth. The notation uses dist(p_{c,i}, Π_c) but does not define dist explicitly. Consider adding a brief definition (e.g., perpendicular distance).","section":"§3.3, Eq. (2)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The spherical direction bias uses G_{ij} = 1 - ω_i^T ω_j, which is the cosine distance. The paper states that β = softplus(b) with learnable b. It would be clearer to state that β is learned per head, not shared across heads.","section":"§3.4, Eq. (6)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The paper mentions that 'Attention uses native PyTorch Scaled Dot-Product Attention (SDPA) with external xFormers and flashAttn disabled for fairness.' It would be helpful to note whether this applies to all models or only to EG360.","section":"§4.1, Comparison settings","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The latency for DA360-S is reported as 517.412 ms, which corresponds to ~1.93 FPS. The abstract states 'a previous ViT-small model runs at less than 2 FPS,' which is consistent. However, the paper should explicitly state that DA360-S is the ViT-small model referred to in the abstract.","section":"§4.2, Table 3","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The LSS loss (least-squares scale alignment) yields very high AbsRel on Metropolis (0.466). The paper attributes this to scale recovery issues. It would be useful to add a brief explanation of why LSS fails on outdoor data.","section":"§4.3, Table 8","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The paper acknowledges limited unlabeled data and temporal incoherence. It would also be relevant to mention that the method relies on two frozen teacher models, which may become outdated as better teachers emerge, and that the distillation pipeline requires caching their predictions offline.","section":"§5, Limitations","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Reference [26] (Metropolis Dataset) is listed as 'Mapillary. Metropolis Dataset, 2024. Accessed: 2025-11-01.' This is a dataset reference; consider providing a more permanent URL or DOI if available.","section":"References","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Hey,\n\nThis paper is a well-engineered system for real-time panorama depth estimation on embedded devices. The main novelty is the adaptive weighting between a panoramic teacher (DAP) and a perspective teacher (MoGe2) based on plane-fitting residuals per cube face. That's a sensible idea: if a face is mostly planar, the panorama teacher's global context is more useful; if it's geometrically complex, the perspective teacher's local detail helps. The spherical direction bias in the cross-level attention is also a nice touch, though it's a small addition.\n\nThe experimental setup is thorough: three zero-shot test sets, comparisons with multiple baselines (some retrained for fairness), and ablation studies for each component. The evaluation is done without scale alignment, which is appropriate. The efficiency numbers are impressive: 6.6M params, 23.5 GFLOPs, 35 FPS on a Jetson Orin NX. That's a real achievement.\n\nThe soft spots are real but not fatal. First, there are no error bars anywhere. All results are single runs. Given the stochasticity in training (data augmentation, batch sampling), this is a concern. Second, the key hyperparameter tau (teacher weighting) is set to 0.3 with no sensitivity analysis. The ablation in Table 7 shows that simple averaging (which corresponds to tau → ∞) already gives strong results, and APPD provides a modest additional improvement. So the sensitivity concern is real but not load-bearing: even if tau is suboptimal, the model still works well. Third, code is not yet released, which limits reproducibility. The authors promise to release it, but that's a promise.\n\nThe paper is honest about its limitations: limited unlabeled data, single-frame only. The citation pattern looks fine.\n\nWho is this for? Researchers working on efficient panorama depth estimation for robotics or VR. It's a solid engineering contribution. The central claim is plausible and supported by the reported numbers. I'd send it to peer review; it deserves a serious referee. The main things to check in review are the sensitivity of tau and the statistical reliability of the results.\n\nBest,\n[Your name]","headline":"Solid engineering paper on efficient panorama depth estimation; the distillation weighting is the main novelty, and the experiments are thorough. The main soft spot is the lack of error bars and sensitivity analysis for the key hyperparameter tau, but the central claim is plausible.","tokens_in":18512,"tokens_out":556,"duration_ms":7075},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"A lightweight panorama depth model runs 35 FPS on a Jetson Orin NX with zero-shot generalization.","keywords":["panorama depth estimation","zero-shot generalization","knowledge distillation","lightweight model","embedded inference","spherical attention","adaptive teacher weighting","equirectangular projection"],"falsifier":"Train EG360 on the same data but with tau set to 0.1 or 1.0 and evaluate zero-shot on Matterport3D and Metropolis; if AbsRel degrades by more than 5% relative on either dataset, the claim that APPD is robust to the choice of tau would be weakened.","tokens_in":19299,"feed_emoji":"🎥","tokens_out":1382,"duration_ms":13252,"temperature":0.7,"pith_summary":"The paper tries to establish that a lightweight panorama depth estimation model can achieve strong zero-shot generalization across indoor and outdoor scenes while running in real time on an embedded platform. It argues that prior lightweight models generalize poorly and prior accurate models are too slow, and that the gap can be closed by combining two ideas: adaptively weighting supervision from a panorama teacher and a perspective teacher based on scene planarity, and fusing cross-level features with a spherical direction bias. The paper reports that the resulting model, EG360, with 6.6M parameters, achieves an average AbsRel of 0.168 across Matterport3D, Stanford2D3D, and Metropolis and runs at 35 FPS on a Jetson Orin NX, whereas a ViT-small model runs at less than 2 FPS. A sympathetic reader would care because it suggests that real-time panoramic depth estimation on mobile robots or VR headsets is feasible without sacrificing generalization.","feed_headline":"Lightweight panorama depth model runs 35 FPS on Jetson Orin NX","feed_subtitle":"Adaptive distillation from panoramic and perspective teachers achieves zero-shot generalization with only 6.6 million parameters.","key_machinery":"The key machinery is the normalized plane-fitting residual r_c for each cubemap face, computed from the perspective teacher's 3D points, which quantifies how planar the face is. This scalar drives the adaptive teacher weighting: w_P = tau/(r_c+tau) for the panoramic teacher and w_C = r_c/(r_c+tau) for the perspective teacher. It is the load-bearing mechanism that decides, per face, which teacher's inverse-depth prediction is more reliable, under the assumption that planar views benefit more from panoramic context. The second mechanism is spherical direction bias in cross-level attention, where the attention logit subtracts $\\beta$ * G_ij with G_ij = 1 - omega_i^T omega_j, encoding spherical proximity and handling the ERP left-right seam naturally.","core_discovery":"EG360 demonstrates that a lightweight panorama depth model (6.593M parameters, 23.492 GFLOPs) can achieve competitive zero-shot accuracy on real-world benchmarks while running at 35 FPS on an embedded Jetson Orin NX, narrowing the gap between expensive foundation models and fast but domain-limited lightweight models. The key enablers are: (1) Adaptive Panorama-Perspective Distillation (APPD), which uses the normalized plane-fitting residual of each cubemap face to dynamically weight the contributions of a panoramic teacher (DAP) and a perspective teacher (MoGe2), assigning more weight to the panoramic teacher for planar views; and (2) Spherical Cross-Level Feature Fusion (SCFF), which applies cross-level attention between the two coarsest encoder feature maps with a spherical direction bias that respects the equirectangular projection's left-right continuity and spherical proximity.","pith_inferences":["The normalized plane-fitting residual could be replaced or augmented with a learned planarity score, which might adapt better to non-planar but structured scenes like furniture clusters.","Because the method relies on two frozen teacher models (DAP and MoGe2), its generalization ceiling is tied to those teachers' blind spots; a future lightweight model that jointly trains student and teachers could potentially surpass both.","The spherical direction bias formulation (1 - omega_i^T omega_j) is essentially a cosine-distance kernel on the sphere; it could be interpreted as a learned temperature-adjusted spherical RBF attention, which might generalize to other spherical modalities like 360-degree audio or LIDAR."],"forward_implications":["Embedded robots and VR headsets can run panoramic depth estimation in real time (35 FPS on Jetson Orin NX) without relying on cloud inference or large models.","The APPD strategy can be applied to other lightweight panorama architectures (UniFuse, BiFuse++, ACDNet) to improve their zero-shot generalization, as shown in Table 9.","The spherical direction bias in SCFF provides a principled way to inject equirectangular geometry into attention mechanisms for any panorama perception task beyond depth.","Training with a simple L1 distillation loss on pseudo labels from complementary teachers outperforms scale-invariant losses (SI, LSS) for zero-shot panorama depth, suggesting that preserving metric scale is important even for unlabeled data."],"supporting_citations":[{"why":"Supplies the lightweight ZipDepth base model that EG360 finetunes as its student backbone.","marker":"[39]"},{"why":"Provides the panoramic teacher DAP, a foundation model for panorama depth that serves as one source of pseudo labels.","marker":"[23]"},{"why":"Provides the perspective teacher MoGe2, a metric-scale depth model that serves as the other source of pseudo labels.","marker":"[47]"},{"why":"Supplies the DA360 training recipe (augmentations, multi-scale gradient loss, ERP padding) that EG360 adapts.","marker":"[18]"},{"why":"Provides the UniFuse lightweight baseline that EG360 compares against and also applies APPD to.","marker":"[17]"}],"fun_headline_variants":[],"cache_read_input_tokens":0,"weakest_assumption_plain":"The distillation hyperparameter tau (set to 0.3) that balances the two teacher weights is chosen experimentally and its sensitivity across different datasets or deployment conditions is not explored; if the optimal tau varies significantly, the claimed robustness may not hold.","fun_headline_variants_meta":{"raw":{"variants":[]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":7.8e-05,"raw_usage":{"cost":7.77322e-05,"is_byok":false,"cost_details":{"upstream_inference_cost":7.77322e-05,"upstream_inference_prompt_cost":7.37282e-05,"upstream_inference_completions_cost":4.004e-06},"total_tokens":1037,"prompt_tokens":1015,"completion_tokens":22,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":0}},"tokens_in":1015,"tokens_out":22,"cost_source":"openrouter_usage","duration_ms":3269,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":0,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-11-13T12:39:48.003549+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Train EG360 on the same data but with tau set to 0.1 or 1.0 and evaluate zero-shot on Matterport3D and Metropolis; if AbsRel degrades by more than 5% relative on either dataset, the claim that APPD is robust to the choice of tau would be weakened.","supporting_citations":[{"cited_title":"Zipdepth: Bringing lightweight zero-shot monoc- ular depth anywhere, on any device","cited_arxiv_id":null,"evidence_quote":"Supplies the lightweight ZipDepth base model that EG360 finetunes as its student backbone."},{"cited_title":"Depth any panoramas: A foundation model for panoramic depth estimation","cited_arxiv_id":null,"evidence_quote":"Provides the panoramic teacher DAP, a foundation model for panorama depth that serves as one source of pseudo labels."},{"cited_title":"Moge-2: Accurate monocular geometry with metric scale and sharp details.Advances in Neural Information Pro- cessing Systems, 38:35928–35959, 2026","cited_arxiv_id":null,"evidence_quote":"Provides the perspective teacher MoGe2, a metric-scale depth model that serves as the other source of pseudo labels."},{"cited_title":"Depth Anything in $360^\\circ$: Towards Scale Invariance in the Wild","cited_arxiv_id":"2512.22819","evidence_quote":"Supplies the DA360 training recipe (augmentations, multi-scale gradient loss, ERP padding) that EG360 adapts."},{"cited_title":"Unifuse: Unidirectional fusion for 360 panorama depth estimation.IEEE Robotics and Automation Letters, 6 (2):1519–1526, 2021","cited_arxiv_id":null,"evidence_quote":"Provides the UniFuse lightweight baseline that EG360 compares against and also applies APPD to."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}