{"id":"8b9ec81d-2abe-4d1e-a012-c64a1e1a5f84","arxiv_id":"2507.17351","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Active learning with a 3D spatial-diversity term cuts annotation cost by over 2x for semantically-aware NeRF training versus random sampling.","lead":"This paper tests whether active learning, which picks the most useful data to label, can cut the cost of adding semantics to NeRF scene models. It proposes a selection strategy that uses the 3D layout of the scene to choose which tiny regions to label, and reports more than twice the annotation savings of random sampling.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The spatial-diversity contribution is not statistically established: ablations are within noise, and Entropy alone already achieves the claimed >2x cost reduction.","rationale":"The reader's weakest_assumption concerned the accuracy of volume-rendered depth in Eq. (5) and whether the spatial term tracks semantics. I agree this is a real fragility, but it is not the most load-bearing issue: even granting accurate depth, the paper's own ablations show the spatial term adds only tenths of a point over Entropy+Feature, and Table I shows plain Entropy achieves essentially the same annotation-cost reduction as the full method. The central empirical claim that the proposed hybrid 'consistently outperforms' Entropy is therefore not established by the reported statistics (3 runs, no significance testing, no code). This does not make the exploration worthless: the finding that active learning, especially entropy-based selection, reduces annotation cost by more than 2x versus random is plausible and visible in the curves, and the superpixel-level granularity study is a useful comparison. The novel term, however, needs stronger evidence. I therefore keep the reader's CONDITIONAL verdict: the authors should add significance testing and confidence intervals on the key comparisons, or soften the claim that the spatial term drives the gains. I set aside the unproven 2-approximation guarantee because it is not needed for the empirical headline, though it should also be addressed.","tokens_in":9851,"tokens_out":7866,"duration_ms":80501,"concrete_test":"Run 10 independent seeds of the Room0 ablation for Entropy+Feature vs Entropy+Feature+Spatial and for Entropy vs Entropy+Spatial, at batches 1-3. Compute paired 95% bootstrap confidence intervals per batch; if any interval for the spatial-term addition includes zero, the claimed benefit of 3D spatial diversity is not supported. As a secondary check, recompute Table I with the same seeds and report confidence intervals for Ours minus Entropy; if this interval includes zero, the 'consistently outperforms Entropy' claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's specific contribution is the 3D spatial diversity term, and the abstract/conclusions claim the hybrid Entropy+Feature+Spatial strategy 'consistently outperforms' Entropy, CoreSet, ViewAL, and Random. The reported evidence does not support a meaningful advantage over plain Entropy. In Table I, the required annotation budget for Ours vs Entropy is 9.50 vs 9.56 (Scene0006), 9.44 vs 9.62 (Scene0030), 9.28 vs 9.56 (Room0), and 9.34 vs 9.43 (Office0): differences of 0.06-0.28 percentage points with no confidence intervals. In the Room0 ablation (Table III), Entropy+Feature+Spatial exceeds Entropy+Feature by only +0.32, +0.39, and +0.02 mIoU at batches 1-3; Entropy+Spatial vs Entropy is +0.08, +0.34, and -0.03. With three runs and standard deviations up to 0.76, these differences are within run-to-run noise. More importantly, Table I shows Entropy alone requires only 9.4-9.6% annotation to reach Random at 20%, i.e., the headline 'more than 2x reduction' is achieved without the proposed spatial term. Thus the paper does not demonstrate that its novel component is necessary for, or even contributes to, the central claim. The load-bearing issue is therefore not depth accuracy per se, but that the measured differences are real; as reported, this is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript investigates active learning for Semantic-NeRF, comparing selection at image and superpixel granularity and several selection strategies (Random, Entropy, CoreSet, ViewAL, and a proposed hybrid). The proposed hybrid adds a 3D spatial diversity term, computed from volume-rendered depths, to entropy and feature diversity, and solves selection via a greedy diversification algorithm claimed to be a 2-approximation. Experiments on Replica and ScanNet scenes over four annotation batches, with three runs each, report that active learning reduces annotation cost by more than 2x relative to random sampling, and the hybrid is claimed to consistently outperform all baselines.","tokens_in":10253,"tokens_out":4550,"duration_ms":46656,"significance":"The paper addresses a timely and underexplored problem with a reasonable experimental design: standard datasets, retraining from scratch each round, and repeated runs with error bars. Its clearest finding—that active learning, including simple entropy, substantially reduces the annotation budget for Semantic-NeRF—is useful to the community. The specific novel contribution, 3D spatial diversity, is not statistically supported, and the theoretical guarantee is under-justified. With strengthened evidence or appropriately narrowed claims, the study could be a worthwhile contribution.","major_comments":[{"comment":"The claim that the proposed Entropy+Feature+Spatial strategy 'consistently outperforms' Entropy is not supported by the reported numbers. In Table I, the required annotation budgets for Ours vs. Entropy differ by only 0.06, 0.18, 0.28, and 0.09 percentage points across the four scenes; in Table III, adding the spatial term to Entropy+Feature improves mIoU by 0.32, 0.39, and 0.02 at batches 1–3 on Room0, while standard deviations across three runs reach 0.76. No significance tests, confidence intervals, or paired analysis are reported, so the differences are within run-to-run noise. This matters because Entropy alone already achieves the headline 'more than 2x reduction' (9.4–9.6% vs. 20%), so the paper does not demonstrate that the proposed component contributes to the central claim.","section":"Section IV-C, Table I, Table III"},{"comment":"The claim that the greedy algorithm in Algorithm 1 provides a 2-approximation for the combined max-min objective in Eq. (2) is asserted without proof. The reference [31] addresses a specific diversification objective, and it is not immediate that the distance redefinition in Eq. (3) preserves the approximation guarantee when unary uncertainty terms are mixed with pairwise distances. Please provide a formal proof or a precise statement of the applicable theorem with conditions.","section":"Section III-A, Eq. (3), Algorithm 1"},{"comment":"The spatial diversity term relies on 3D coordinates C(x) obtained from volume-rendered depths of the current Semantic-NeRF, which is trained with very sparse labels in early batches. The paper does not evaluate depth accuracy or its influence on selection quality. Since the effectiveness of the spatial term is a central claim, this omission leaves the small observed gains unexplained: they could stem from geometry errors rather than from meaningful 3D diversity.","section":"Section III-B, Eq. (5)–(6)"}],"minor_comments":[{"comment":"The phrase 'The most close work' should be 'The closest work.'","section":"Section I"},{"comment":"The normalization of d_f and d_s to [0,1] is not described; please specify how the scaling is computed (e.g., per-batch min-max normalization) and whether it affects the metric property claimed in Eq. (6).","section":"Section III-B"},{"comment":"Table I reports interpolated budgets but does not state the interpolation method or provide confidence bounds; please add this information.","section":"Section IV-C, Table I"},{"comment":"The ablation table would be clearer with a column for batch 0 and with explicit differences between variants rather than only absolute means and standard deviations.","section":"Section IV-E, Table III"},{"comment":"The supplementary plots in Fig. 6 include error bars only in the text description; please ensure the captions and figures clearly indicate the number of runs and the meaning of the bars.","section":"Appendix B"}],"recommendation":"major_revision","confidential_remarks":"The paper's empirical claims are broader than the evidence. The authors should either add more repetitions or significance tests, or revise the abstract and conclusions to state that simple entropy already provides most of the benefit. The 2-approximation guarantee needs a formal proof or a precise reference. The study is otherwise within the journal's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's the short version: this is a genuinely useful empirical study of active learning for semantic NeRF, but the paper's specific selling point--the 3D spatial diversity term--is not supported by the numbers. The broad claim that active learning cuts annotation cost >2x holds up. The narrower claim that their hybrid strategy 'consistently outperforms' Entropy does not.\n\nWhat's new: this is the first AL study for semantic NeRF that I know of. The system is standard: entropy for uncertainty, feature (logit) distance for diversity, plus a 3D spatial distance from volume-rendered depth. They compare superpixel vs image-level sampling, run three seeds, retrain from scratch each batch, and use Replica and ScanNet. That's solid experimental hygiene, and the superpixel-level result (better than image-level for both Entropy and CoreSet) is a useful, likely reproducible finding.\n\nThe soft spots are real. Table III shows adding the spatial term to Entropy changes mIoU by +0.08, +0.34, -0.03 over the three batches; adding it to Entropy+Feature gives +0.32, +0.39, +0.02. Standard deviations are up to 0.76. These are noise, not signal. Meanwhile Table I shows Entropy by itself needs 9.4-9.6% annotation to match Random at 20%--so the headline 'more than 2x reduction' is achieved without their proposed component. No significance tests are reported anywhere. The 'consistently outperforms' language is an overreach.\n\nTwo smaller issues. The 2-approximation guarantee in Sec III-A is asserted, not proven; the greedy algorithm may work, but you can't take that on faith. And the spatial term relies on depth from a NeRF trained on very sparse labels, so its reliability is questionable--though that's secondary, since the empirical gain from the term isn't clear either way.\n\nVerdict: the empirical core of the paper is sound and worth publishing after revision. A referee can reasonably ask for significance testing or at least confidence intervals on the budget comparisons, and for the 2-approx claim to be either proved or dropped. The authors have done the hard work of a systematic comparison; they just need to match the claims to the evidence.\n\nRecommended action: send it to peer review. It deserves a serious referee, and a major revision would make it a decent paper.","headline":"Useful empirical study of active learning for semantic NeRF, but the novel spatial-diversity term is not statistically supported; Entropy alone already delivers the headline >2x cost reduction.","tokens_in":10730,"tokens_out":3263,"would_cite":false,"duration_ms":31567,"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":"Active learning cuts semantic NeRF annotation cost by half","keywords":["active learning","semantic neural radiance field","label-efficient training","superpixel selection","uncertainty sampling","diversity sampling","3D spatial diversity"],"falsifier":"Retrain the proposed selection on a scene while replacing the predicted depth in Equation 5 with ground-truth depth from the RGB-D sensors used by Replica and ScanNet; if the predicted-depth version does not lose noticeable mIoU relative to the ground-truth-depth version, or if the advantage over Entropy shrinks, then the 3D spatial term is not doing the work claimed.","tokens_in":9686,"feed_emoji":"🎯","tokens_out":3917,"duration_ms":36670,"temperature":0.7,"pith_summary":"This paper asks whether active learning can make training a semantically-aware neural radiance field affordable, where the model must learn both 3D geometry and per-point semantic labels. The authors claim it can: their selection strategy reaches the same segmentation quality as random sampling while labelling roughly half as many pixels. The strategy picks small image regions, or superpixels, that the model is most uncertain about while also spreading selections across the scene, using both 2D features and 3D positions recovered from the model's own depth prediction. If this holds, semantic 3D scene understanding becomes far cheaper to annotate.","feed_headline":"Active learning cuts semantic NeRF annotation cost by half","feed_subtitle":"Choosing uncertain, well-spread superpixels beats random pixel labeling for 3D scene understanding.","key_machinery":"The load-bearing object is the max-min diversification objective with a combined distance function $d'(x, y) = \\frac{1}{2}(u(x) + u(y)) + d(x, y)$, where $u$ is prediction entropy and $d(x, y) = d_f(x, y) + d_s(x, y)$ is the sum of feature L2 distance and 3D spatial L2 distance between superpixels. The 3D coordinates for the spatial term come from volume-rendered depth (Equation 5): each pixel's depth is the expectation along its ray, giving a 3D point per superpixel. Selection is solved greedily with a 2-approximation algorithm (Algorithm 1), keeping complexity $O(n_b \\cdot n_u \\cdot f_{\\text{dim}})$, comparable to CoreSet's k-center greedy.","core_discovery":"On the paper's terms, the central discovery is that a hybrid active-learning criterion combining prediction entropy, feature diversity, and a newly added 3D spatial-diversity term consistently outperforms Entropy, CoreSet, ViewAL, and random sampling for training Semantic-NeRF on Replica and ScanNet indoor scenes across annotation budgets. Measured by the budget needed to match random sampling at 20% annotation, the proposed method needs only about 9.3 to 9.5 percent of regions, a more than two-fold reduction in annotation cost. The paper also reports that superpixel-level selection is more cost-effective than image-level selection, and that ViewAL's viewpoint-consistency criterion is ineffective for NeRF because NeRF predictions are multi-view consistent by construction.","pith_inferences":["The 3D spatial-diversity term could generalize to other NeRF-based label-efficient tasks, such as panoptic segmentation or open-vocabulary segmentation, where geometry is also available from the rendering pipeline.","Because the spatial term depends on predicted depth, a testable extension is to weight the spatial term by estimated depth confidence or to switch to ground-truth depth when available; the paper does not ablate this dependence.","The same max-min diversification framing could be applied to other 3D scene representations, for example 3D Gaussian splatting, without re-deriving the greedy solver.","The two-fold reduction is measured against random sampling at a fixed 20% budget on four indoor scenes; the claim generalizes only if the trend holds across more scenes and tighter budgets."],"forward_implications":["Annotation budgets for semantic NeRF can be cut by more than half while keeping the same segmentation quality.","Region-level (superpixel) active selection is more cost-effective and more stable than image-level selection for Semantic-NeRF.","Hybrid uncertainty-plus-diversity selection beats either criterion alone across scenes with different characteristics.","Viewpoint-consistency cues such as those in ViewAL add little for NeRF because the semantic branch is already viewpoint invariant."],"supporting_citations":[{"why":"Defines the Semantic-NeRF base model that the active learning loop trains and evaluates.","marker":"[5]"},{"why":"ViewAL is the closest prior work, used as a baseline and argued to be ineffective for NeRF because of built-in multi-view consistency.","marker":"[7]"},{"why":"CoreSet is the diversity-based baseline and the source of the k-center greedy selection idea that the proposed method extends.","marker":"[15]"},{"why":"Shannon entropy supplies the uncertainty measure u(x) used in the selection objective.","marker":"[29]"},{"why":"The result diversification framework provides the max-min objective and the 2-approximation greedy algorithm used to solve it.","marker":"[31]"},{"why":"Replica is one of the two indoor scene datasets used for evaluation.","marker":"[34]"},{"why":"ScanNet is the other evaluation dataset, covering varied indoor scene categories.","marker":"[35]"},{"why":"SEEDS is the superpixel algorithm that defines the region-level selection granularity.","marker":"[37]"}],"fun_headline_variants":["3D-aware active learning trims semantic NeRF labels by over 2x","Superpixel active learning halves semantic NeRF annotation cost","3D diversity boosts active learning for semantic NeRF labeling","Active learning needs just 9.5% labels to match 20% random"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 3D spatial-diversity term assumes the volume-rendered depth from the current Semantic-NeRF model is accurate enough to place superpixels at their true 3D locations, even in early active-learning rounds when the model was trained on only five percent of labelled regions.","fun_headline_variants_meta":{"raw":{"variants":["3D-aware active learning trims semantic NeRF labels by over 2x","Superpixel active learning halves semantic NeRF annotation cost","3D diversity boosts active learning for semantic NeRF labeling","Active learning needs just 9.5% labels to match 20% random"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000631,"raw_usage":{"total_tokens":2858,"prompt_tokens":836,"completion_tokens":2022,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":452,"completion_tokens_details":{"reasoning_tokens":1945}},"tokens_in":452,"tokens_out":2022,"duration_ms":16175,"temperature":1.0,"reasoning_tokens":1945,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:49:44.318590+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the proposed selection on a scene while replacing the predicted depth in Equation 5 with ground-truth depth from the RGB-D sensors used by Replica and ScanNet; if the predicted-depth version does not lose noticeable mIoU relative to the ground-truth-depth version, or if the advantage over Entropy shrinks, then the 3D spatial term is not doing the work claimed.","supporting_citations":[{"cited_title":"In-place scene labelling and understanding with implicit scene representation,","cited_arxiv_id":null,"evidence_quote":"Defines the Semantic-NeRF base model that the active learning loop trains and evaluates."},{"cited_title":"Viewal: Active learning with viewpoint entropy for semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"ViewAL is the closest prior work, used as a baseline and argued to be ineffective for NeRF because of built-in multi-view consistency."},{"cited_title":"A mathematical theory of communication,","cited_arxiv_id":null,"evidence_quote":"Shannon entropy supplies the uncertainty measure u(x) used in the selection objective."},{"cited_title":"An axiomatic approach for result diversification,","cited_arxiv_id":null,"evidence_quote":"The result diversification framework provides the max-min objective and the 2-approximation greedy algorithm used to solve it."},{"cited_title":"The replica dataset: A digital replica of indoor spaces,","cited_arxiv_id":null,"evidence_quote":"Replica is one of the two indoor scene datasets used for evaluation."},{"cited_title":"Scannet: Richly-annotated 3d reconstructions of indoor scenes,","cited_arxiv_id":null,"evidence_quote":"ScanNet is the other evaluation dataset, covering varied indoor scene categories."},{"cited_title":"Seeds: Superpixels extracted via energy-driven sampling,","cited_arxiv_id":null,"evidence_quote":"SEEDS is the superpixel algorithm that defines the region-level selection granularity."}],"review_version":1}