{"id":"15f6ad13-f1b8-4a6f-b8da-d9ad61aa3761","arxiv_id":"1908.01379","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An RGB-superpixel-guided depth sampling and reconstruction method with one sample per segment plus bilateral filtering outperforms random and grid sampling on indoor and outdoor benchmarks.","lead":"This paper uses RGB image superpixels to decide where a depth sensor should point, taking one depth measurement per segment and then smoothing the result. The authors report that this reduces the number of depth samples needed for accurate reconstruction by roughly three to four times compared with random or grid sampling.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 3-4x claim conflates sampling and reconstruction: Table 1 compares the full proposed pipeline against baselines given uniform random samples, so the factor is not isolated to the reconstruction.","rationale":"The reader's weakest assumption was that RGB superpixel boundaries must match depth discontinuities. That is a legitimate limitation and is acknowledged in the paper, but the central claim as stated in the abstract is a quantitative superiority claim about 'our proposed reconstruction' relative to image-guided depth completion algorithms. The more load-bearing unsecured condition is that the comparison is apples-to-apples: the proposed method gets image-adaptive sample locations while the baselines get uniform random locations. The paper's own sampling-only experiment shows the sampling pattern alone helps other methods, so without a matched-sampling ablation the factor of 3-4 cannot be attributed to the reconstruction component. I do not think this invalidates the paper; the system-level contribution may still be valuable, and the reader's CONDITIONAL verdict is appropriate. The condition should explicitly require either (a) a matched-sampling comparison isolating the reconstruction, or (b) a reframing of the claim so that 'sampling and reconstruction system' is what is being compared. My disagreement with the reader is therefore about which assumption is the weakest link in the central claim, not about the overall verdict.","tokens_in":9544,"tokens_out":7920,"duration_ms":89022,"concrete_test":"Recompute Table 1 (Synthia obstacles, full Synthia, and NYU-Depth-v2) with all baseline reconstruction methods (bilinear interpolation, bilateral solver, IP-Basic, L1diag) using exactly the proposed superpixel center-of-mass sample locations at matched sample counts, and separately run the proposed zero-order plus bilateral reconstruction on uniform random samples at the same counts. If baseline methods with the proposed samples reach the target RMSE within roughly 1.5x fewer samples than Ours, or if Ours with random samples degrades to baseline-level RMSE, then the headline factor is a sampling-pattern effect and the reconstruction-specific claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing unsecured step is the attribution of the headline 3-4x sample reduction to the proposed reconstruction. In Sec. 5.1 the authors state that 'for all other examined methods we simulate uniform random depth samples at different sparsity levels', while the proposed method samples once at each RGB superpixel center of mass and reconstructs from those same superpixels. Table 1 therefore measures a combined sampling-plus-reconstruction system against reconstruction-only baselines fed random samples. Since the paper itself shows in Sec. 5.1/Fig. 14 that the proposed sampling pattern alone improves other completion methods, the 3-4x factor in Table 1 may be contributed mostly by the informative sample locations rather than by the zero-order plus bilateral reconstruction. The abstract's wording ('our proposed reconstruction achieves... reducing the required sampling rate by a factor of 3-4') is not supported by an ablation that isolates the reconstruction. A related overstatement is that the 3-4x number appears only on the Synthia obstacles subset; full-scene results in Fig. 12a show about 30% RMSE improvement at fixed density, and NYU Table 2 shows no factor-of-3-4 gain at fixed sample count. The acknowledged camouflage limitation (Sec. 4.3.1) is real, but the matched-sampling question is more directly load-bearing for the central claim.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an adaptive depth sampling and reconstruction framework guided by RGB superpixels. The authors first introduce a piecewise planar depth model, estimate on Synthia and NYU-Depth-v2 that 20-60 planar regions approximate typical indoor and outdoor depth maps, and derive an optimal sampling ratio near 1/1200 from the 3N lower bound. The proposed algorithm samples once at the center of mass of each SLIC superpixel and reconstructs by zero-order assignment followed by bilateral filtering in the log domain. Experiments compare against several depth completion baselines on Synthia and NYU, reporting lower RMSE and, on a Synthia obstacle subset, a 3-4x reduction in required sampling density. A single-pixel mechanical depth sampler demonstrates the concept in a laboratory setting.","tokens_in":9827,"tokens_out":5109,"duration_ms":52135,"significance":"If the central claims hold, the work offers a simple, non-learning, image-guided sampling strategy that is relevant to emerging solid-state LiDAR systems, and it provides a falsifiable prediction about the number of planar structures needed for depth approximation. The paper's explicit strengths are the breadth of the evaluation (multiple reconstruction baselines, indoor/outdoor data, a resolution analysis via MTF, and a physical prototype) and the consistent qualitative gains shown across settings. The main weaknesses are that the headline sample-reduction factor is not cleanly attributed to the reconstruction rather than the sampling pattern, and several implementation details needed for reproducibility are missing. The contribution is useful and appears technically sound in its core comparison, but the claims need to be re-scoped or supported by additional ablations before publication.","major_comments":[{"comment":"The statement that 'our proposed reconstruction achieves state-of-the-art results... reducing the required sampling rate by a factor of 3-4' attributes the gain to the reconstruction, but the comparison in Table 1 varies both the sampling pattern and the reconstruction jointly. For all baselines the paper simulates uniform random samples, while the proposed method samples once per RGB superpixel at the center of mass; Sec. 5.1 and Fig. 14 show that this sampling pattern alone improves other completion methods. Consequently, the 3-4x factor in Table 1 is not evidence that the zero-order-plus-bilateral reconstruction is responsible for the gain. Please provide an ablation that fixes the sampling pattern (e.g., our sampling with the baselines' reconstructions, or our reconstruction with random/grid samples) and report sample-requirement curves for that ablation.","section":"Sec. 5.1 (Table 1) and Abstract"},{"comment":"The procedure used to compute the piecewise planar approximations is not described. The paper reports average N, δ, and ε for Synthia and NYU, but without specifying how the partition E and plane parameters (a_i, b_i, c_i) are estimated, how the validity set V is determined, and what optimization or segmentation algorithm is used, these numbers and the derived 1/1200 optimal sampling ratio cannot be reproduced. This is load-bearing because Eq. (3) and the model-validity claim rest on these estimates.","section":"Sec. 3.1"},{"comment":"The bilateral filter parameters are never given; the text only says they are 'fixed for a given number of samples n and type of scene'. Since the reconstruction is central to the claimed improvements, the spatial and range standard deviations (and any iteration count) must be reported for each experiment. The SLIC compactness parameter in step S.1 is also unspecified. Without these values, the method cannot be reimplemented or compared fairly.","section":"Sec. 4.2 (R.3) and Sec. 5"},{"comment":"The 3-4x sample reduction is demonstrated only on the Synthia obstacles subset. On the full Synthia scene the improvement at fixed density is about 30% RMSE (Fig. 12a), and on NYU-Depth-v2 (Table 2) the gain at a fixed budget of 200 samples is modest. The abstract and introduction state the 3-4x factor without this scope restriction. Please qualify the claim to the obstacles evaluation or provide equivalent density-vs-RMSE curves for the full scenes, including the sample counts required for a given RMSE.","section":"Sec. 5.1 (Figs. 12a/12b, Table 1)"}],"minor_comments":[{"comment":"In the sentence 'An example of output of a our algorithm', 'a our' should be 'our'.","section":"Sec. 1"},{"comment":"The notation 'Number−of−Samples' is awkward; consider using a mathematical symbol such as n_min.","section":"Sec. 3.1, Eq. (3)"},{"comment":"The caption says 'percentage of approximated planes N per image' but the correct term is 'number of approximated planes'.","section":"Fig. 3 caption"},{"comment":"The text says the desired number of SPs is set to n, but it should clarify how the actual sample count relates to n when a center of mass falls outside a non-convex superpixel.","section":"Sec. 4.2, step S.1"},{"comment":"Uniform random sampling is stochastic, so the baseline curves in Figs. 12 and Table 1 should include error bars or be averaged over multiple trials to establish that the reported differences are not due to sampling noise.","section":"Sec. 5.1"},{"comment":"Please specify the training data and preprocessing for the deep-learning baselines, since a fair comparison requires that they are evaluated under conditions consistent with the NYU test set.","section":"Sec. 5.2, Table 2"}],"recommendation":"major_revision","confidential_remarks":"The core idea is timely and the experiments are broad, but the headline claim needs either an ablation that isolates the reconstruction contribution or a re-scoped statement. The missing fitting and filtering details are straightforward to supply and are important for reproducibility. I see no grounds for rejection, but the manuscript is not ready in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, this paper is a decent engineering contribution, but read the 3-4x claim with caution. The core idea—using RGB superpixels to decide where to sample depth, one sample per segment, then a log-domain bilateral filter to clean up zero-order reconstruction—is new and works reasonably well. The piecewise planar complexity study (20-60 regions, 1/1200 optimal ratio) is a useful back-of-the-envelope baseline, though the fitting procedure is not described. Credit where due: they show consistent improvements over several classical and learned baselines, build a physical single-pixel sampler, and add an MTF analysis. That's real work.\n\nThe soft spots are mostly about attribution and evidence. The abstract claims the reconstruction reduces sampling rate 3-4x, but Table 1 compares the full sampling-plus-reconstruction system against baselines fed uniform random samples. Their own Fig. 14 shows the sampling pattern alone helps other completion methods, so the headline factor likely comes largely from sampling, not from the proposed reconstruction. That is a load-bearing issue for how the result is stated. Full-scene Synthia results show about 30% RMSE improvement at fixed density, not 3-4x, and NYU Table 2 shows a modest gain at 200 samples. The 3-4x number appears only on the obstacles subset, which is the most favorable case. Also, bilateral filter parameters are not given, no error bars, and the outdoor evaluation is synthetic. None of this makes the method useless, but the central claim as written is overreach.\n\nThe camouflage failure (Sec 4.3.1) is honestly discussed and is a real limitation, but the matched-sampling question is more directly relevant to the headline.\n\nVerdict: this deserves a serious referee. It's not a desk reject. But the authors should be asked to add an ablation that separates sampling from reconstruction, disclose the missing parameters, and revise the abstract. For someone working on adaptive depth sensing, this is worth a read. I'd probably cite it if I did a survey, but it's not going into my immediate work.","headline":"Solid adaptive depth sampling paper, but the headline 3-4x claim conflates sampling with reconstruction and needs an ablation before it should be accepted as stated.","tokens_in":10344,"tokens_out":2901,"would_cite":false,"duration_ms":26006,"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":"The paper claims that sampling one depth value per RGB superpixel, then smoothing with a log-domain bilateral filter, reconstructs depth with 3-4 times fewer samples than state-of-the-art image-guided depth completion.","keywords":["adaptive depth sampling","depth completion","superpixels","piecewise planar model","bilateral filter","RGB-guided sampling","LiDAR","sparse depth reconstruction"],"falsifier":"Measure the required sample count on a benchmark deliberately filled with camouflaged objects, where depth changes occur with no visible RGB edge; if the 3-4x sample savings disappears or reverses, the central claim is bounded to scenes with visible color boundaries. Alternatively, fit the piecewise planar model to highly curved scenes such as vegetation or cables and check whether the number of planes N grows beyond the claimed 20-60 range.","tokens_in":9359,"feed_emoji":"📷","tokens_out":4564,"duration_ms":45859,"temperature":0.7,"pith_summary":"The paper asks how few depth measurements a navigation system truly needs when a color camera can say where to look. It argues that depth scenes are mostly piecewise planar, that 20-60 planar regions describe a typical frame, and therefore that an optimal sampler would need only about one depth reading per 1200 pixels. The proposed practical scheme does not reach that bound, but it cuts the required sampling rate by a factor of 3-4 relative to existing image-guided depth completion methods by taking one measurement per superpixel center of mass and filtering in the logarithmic domain. A physical single-pixel scanning prototype demonstrates the idea outside simulation.","feed_headline":"Depth sampler needs 3-4x fewer probes when guided by color","feed_subtitle":"Adaptive scanning guided by an RGB camera reconstructs depth at a fraction of the sample budget.","key_machinery":"The central object is the superpixel map, computed by SLIC, which simultaneously defines where samples are taken (superpixel centers of mass) and the initial zero-order reconstruction (each superpixel takes its sample's depth value). The second mechanism is a bilateral filter applied in the log domain, which smooths staircase artifacts on large planar walls and roads while preserving real depth discontinuities. These two pieces are justified by a piecewise planar model of depth scenes, validated statistically on dense indoor and outdoor data.","core_discovery":"Depth sampling and reconstruction can be driven by a superpixel segmentation of the RGB image. The method's sampling places one depth probe at the center of mass of each superpixel; reconstruction assigns that depth to the entire superpixel, then applies a bilateral filter to the logarithm of the depth-plus-one map and exponentiates back. The paper reports that this simple pipeline achieves state-of-the-art RMSE on indoor and outdoor benchmarks (NYU-Depth-v2 and Synthia), needing 3-4 times fewer samples than competing methods for a given accuracy, especially on small obstacles. Under the paper's piecewise planar model, three samples per plane give a theoretical lower bound of roughly 56-200 samples per frame, about 0.08 percent of pixels, which marks how much room remains below the achieved sampling rates.","pith_inferences":["If one quantified per frame how well superpixel boundaries match depth discontinuities, the 3-4x savings could be predicted in advance, and the sampler could gracefully fall back to grid sampling on poorly matched frames—a testable extension the paper does not run.","The log-domain bilateral smoothing approximates relative depth error; filtering in inverse-depth or disparity space might yield similar or better accuracy, especially for close objects.","The 20-60 plane model suggests a compressed depth representation for transmission: rather than dense depth, send plane parameters plus a superpixel segmentation, which could approach the theoretical sampling bound.","The paper's assumptions of clean range readings are optimistic; under realistic LiDAR noise the advantage may shrink, so a study of sample-noise sensitivity would bound the method's practical gain."],"forward_implications":["Solid-state LiDAR with programmable scanning can use an RGB camera to cut scan time and energy by a factor of 3-4 at equal reconstruction accuracy.","Because the sampling pattern alone improves other completion methods, the superpixel center-of-mass pattern is a generic front-end that can be paired with any reconstruction algorithm.","The 1/1200 optimal sampling ratio gives a concrete target for future adaptive samplers, indicating a large gap between current practical methods and the theoretical lower bound.","Reliable reconstruction of thin obstacles like poles and signs at low sampling budgets improves detection of small hazards in autonomous navigation."],"supporting_citations":[{"why":"SLIC superpixel segmentation supplies the sampling grid and the initial zero-order reconstruction regions.","marker":"[1]"},{"why":"The fast bilateral solver is the main state-of-the-art baseline that the paper's reconstruction outperforms.","marker":"[6]"},{"why":"IP-Basic is a classical depth completion baseline used in the quantitative and qualitative comparisons.","marker":"[19]"},{"why":"L1diag is a sparse depth completion baseline against which the sample savings are measured.","marker":"[24]"},{"why":"The sparse-to-dense deep learning method is a state-of-the-art baseline on NYU-Depth-v2.","marker":"[26]"},{"why":"Earlier adaptive sampling work by Liu et al. is compared as a sampling-pattern baseline.","marker":"[22]"},{"why":"The bilateral filter is the edge-preserving smoothing operator used in the log-domain reconstruction step.","marker":"[34]"},{"why":"The Synthia dataset provides dense synthetic outdoor depth used for model validation and evaluation.","marker":"[30]"},{"why":"NYU-Depth-v2 provides the indoor benchmark with aligned RGB and dense depth for evaluation.","marker":"[32]"},{"why":"Structured-forest edge detection computes RGB boundaries for the statistical analysis of RGB-depth edge correlation.","marker":"[13]"}],"fun_headline_variants":["RGB guidance cuts depth samples 3-4x","Image-guided depth sampling needs 3-4x fewer probes","Superpixel-guided depth sampling uses 3-4x fewer probes","Image-driven depth sampling cuts sample count by 3-4x","Adaptive RGB-guided depth sampling: 3-4x fewer probes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that RGB superpixel boundaries align closely with true depth discontinuities and that one depth reading at a superpixel's center of mass represents the whole segment; the paper itself shows a camouflaged object where this fails and the method degenerates to ordinary grid sampling.","fun_headline_variants_meta":{"raw":{"variants":["RGB guidance cuts depth samples 3-4x","Image-guided depth sampling needs 3-4x fewer probes","Superpixel-guided depth sampling uses 3-4x fewer probes","Image-driven depth sampling cuts sample count by 3-4x","Adaptive RGB-guided depth sampling: 3-4x fewer probes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000582,"raw_usage":{"total_tokens":2731,"prompt_tokens":928,"completion_tokens":1803,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":1714}},"tokens_in":544,"tokens_out":1803,"duration_ms":13168,"temperature":1.0,"reasoning_tokens":1714,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:14:24.509973+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the required sample count on a benchmark deliberately filled with camouflaged objects, where depth changes occur with no visible RGB edge; if the 3-4x sample savings disappears or reverses, the central claim is bounded to scenes with visible color boundaries. Alternatively, fit the piecewise planar model to highly curved scenes such as vegetation or cables and check whether the number of planes N grows beyond the claimed 20-60 range.","supporting_citations":[{"cited_title":"Achanta, A","cited_arxiv_id":null,"evidence_quote":"SLIC superpixel segmentation supplies the sampling grid and the initial zero-order reconstruction regions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The fast bilateral solver is the main state-of-the-art baseline that the paper's reconstruction outperforms."},{"cited_title":"In Defense of Classical Image Processing: Fast Depth Completion on the CPU","cited_arxiv_id":"1802.00036","evidence_quote":"IP-Basic is a classical depth completion baseline used in the quantitative and qualitative comparisons."},{"cited_title":"Sparse Depth Sensing for Resource-Constrained Robots","cited_arxiv_id":"1703.01398","evidence_quote":"L1diag is a sparse depth completion baseline against which the sample savings are measured."},{"cited_title":"Ma and S","cited_arxiv_id":null,"evidence_quote":"The sparse-to-dense deep learning method is a state-of-the-art baseline on NYU-Depth-v2."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Earlier adaptive sampling work by Liu et al. is compared as a sampling-pattern baseline."},{"cited_title":"Tomasi and R","cited_arxiv_id":null,"evidence_quote":"The bilateral filter is the edge-preserving smoothing operator used in the log-domain reconstruction step."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The Synthia dataset provides dense synthetic outdoor depth used for model validation and evaluation."},{"cited_title":"Silberman, D","cited_arxiv_id":null,"evidence_quote":"NYU-Depth-v2 provides the indoor benchmark with aligned RGB and dense depth for evaluation."},{"cited_title":"Doll ´ar and C","cited_arxiv_id":null,"evidence_quote":"Structured-forest edge detection computes RGB boundaries for the statistical analysis of RGB-depth edge correlation."}],"review_version":1}