{"id":"22879d22-4c30-4146-b7c7-c804182c570a","arxiv_id":"2505.19863","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"FruitNeRF++ counts fruits in orchards by learning 3D instance embeddings with a contrastively trained neural instance field and clustering them with a shape-agnostic HDBSCAN.","lead":"The paper trains a neural radiance field to learn per-fruit instance identities from 2D segmentation masks, then counts fruits by clustering the resulting 3D point cloud. It aims to replace fruit-specific template matching with a shape-agnostic pipeline that works across multiple fruit types.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'fully fruit-agnostic' claim is contradicted by the method's own data: plum and mango require fruit-specific prompt additions and their recall drops to 0.341/0.392 with SAM masks, showing counting accuracy is governed by per-fruit segmentation quality, not by the shape-agnostic NeRF/clustering…","rationale":"The paper's central claim is that the counting pipeline is fully agnostic to fruit type, needing no shape priors. I inspected whether that claim is supported by the reported experiments. The NeRF and clustering stages are indeed template-free, but the input instance masks come from Grounded-SAM/Detic, and the authors modify the text prompt for plum and mango (Sec. IV-A) because the default prompts yield poor masks. This is a fruit-specific intervention inside the pipeline, and it is not reflected in the Sec. VII claim. Table II quantifies the consequence: SAM-mask recall is 0.341 for plum and 0.392 for mango, versus 0.759/0.557 with ground-truth masks; other fruits stay near 0.94+. Table I shows these failures track mask IoU (0.382 and 0.588). Thus the end-to-end method is accurate only when the foundation model happens to segment the fruit type well; it is not demonstrably fruit-agnostic. A concrete test is to replace the per-fruit prompts with a single generic prompt and see whether plum/mango degrade further or recover. If they degrade, the claim fails; if they recover, the prompt modification was load-bearing. I also note the abstract's 'compares favorably' is not supported by the FUJI result (0.765 vs 0.881 for Gené-Mola), though the authors attribute this to pose noise. Given these empirical gaps, the reader's CONDITIONAL verdict is appropriate, but the condition should explicitly require evidence of fruit-agnostic input handling, not just fixing mask-splitting within a fruit. The paper is otherwise clearly written, releases code, and provides a useful extension of FruitNeRF; these points do not offset the unverified central claim but do justify a conditional rather than harsh rejection.","tokens_in":13002,"tokens_out":9416,"duration_ms":95197,"concrete_test":"Run the full synthetic evaluation (Table II) with Grounded-SAM masks obtained using the single text prompt 'fruit' for all six fruit types, keeping D=32, tau=0.2, lambda_c=1, lambda_e=1, and all other pipeline settings unchanged. Compare per-fruit F1 against Table II's SAM-mask column. If plum and mango F1 remain below 0.55 while the other four fruits stay above 0.9, the 'fully fruit-agnostic' claim is contradicted; if the F1 spread narrows, the prompt engineering, not the counting method, was the source of the reported variation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section IV-A, the authors state that for plum and mango they appended 'apple' to the Grounded-SAM text prompt because using only 'mango' or 'plum' produced poor masks. This is an explicit fruit-specific intervention, which contradicts the conclusion (Sec. VII) that the approach is 'fully agnostic to the type of fruit, regarding both the neural radiance field and the clustering.' More importantly, Table II shows that with SAM-generated masks, plum recall falls to 0.341 and mango recall to 0.392 (from 0.759 and 0.557 with GT masks), while apple/lemon/pear/peach remain near 0.94-1.0. This large per-fruit variation tracks the mask IoU in Table I (plum 0.382, mango 0.588) rather than any fruit-independent property of the NeRF/clustering stage. Thus the pipeline's counting accuracy is dominated by the quality of the upstream instance masks, which are not fruit-agnostic in practice. The paper's central claim therefore does not hold end-to-end: the method is a shape-agnostic clustering scheme conditioned on a fruit-dependent segmentation module, and its generalization across fruit types is not demonstrated. Additionally, the abstract's 'compares favorably' is contradicted by the FUJI result, where FruitNeRF++ (F1=0.765) trails Gené-Mola et al. (F1=0.881).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"FruitNeRF++ proposes a fruit-counting pipeline that combines a neural radiance field with a contrastively trained neural instance field. The method takes 2D instance masks from foundation models (Grounded-SAM, Detic) and lifts them into 3D via contrastive learning, then produces a fruit count by clustering instance-embedded fruit points with HDBSCAN using a Euclidean-plus-cosine distance. The paper evaluates on a synthetic multi-fruit dataset (apple, plum, lemon, pear, peach, mango) and on the real FUJI apple dataset, reporting F1-scores for different mask sources and comparing with FruitNeRF and Gené-Mola et al. The central claim is that the method is fully fruit-agnostic in both the neural field and the clustering stages.","tokens_in":13335,"tokens_out":4767,"duration_ms":43943,"significance":"If the method delivered on its claims, it would be a valuable step toward practical multi-fruit counting without fruit-specific shape templates. The contrastive instance-field formulation with local hard negatives for small objects is a plausible and potentially useful contribution, and the paper ships a project page and promises code, which increases reproducibility. However, the strength of the claims currently exceeds the evidence: the 'fully fruit-agnostic' conclusion is contradicted by the authors' own fruit-specific prompt adjustments and by the large mask-quality-dependent performance gaps; the hyperparameter protocol is inconsistent; and the real-world comparison does not support the abstract's 'compares favorably' wording. The core idea is promising, but the evaluation and claims need substantial rework before the manuscript meets the standard of a journal publication.","major_comments":[{"comment":"The claim in Sec. VII that the approach is 'fully agnostic to the type of fruit, regarding both the neural radiance field and the clustering' is contradicted by the method description in Sec. IV-A, where the authors state that for plum and mango they added 'apple' to the Grounded-SAM prompt because the fruit-specific prompt alone produced poor masks. This is an explicit fruit-specific intervention in the pipeline. Moreover, Table II shows that with SAM-generated masks, plum recall drops to 0.341 and mango recall to 0.392 (from 0.759 and 0.557 with GT masks), while apple, lemon, pear, and peach remain near 0.94–1.0. This pattern tracks the segmentation IoU in Table I (plum 0.382, mango 0.588) rather than any property of the NeRF or clustering stage. The end-to-end method is therefore not fruit-agnostic; it is a shape-agnostic clustering scheme whose counting accuracy is dominated by a fruit-dependent segmentation module. Please rephrase the central claim to describe the NeRF and clustering stages as shape-agnostic, and add a quantitative analysis of how mask quality propagates to counting error.","section":"Sec. IV-A, Sec. VII, Table II"},{"comment":"There is a direct inconsistency in the hyperparameters. Sec. IV-E states that the Euclidean distance weight in Eq. (9) is set to lambda_e = 5, while Sec. V begins by saying 'We used the same parameters for all experiments' and lists lambda_e = 1. Later in Sec. V, the FUJI experiment is reported with tau = 0.35 and lambda_e = 5, again different from the synthetic experiments. Given that Fig. 5 shows the results are sensitive to these parameters, this inconsistency suggests that hyperparameters were selected per dataset without a held-out validation split. The paper reports no error bars and no validation protocol, making it impossible to assess whether the reported numbers generalize. Please clarify the exact hyperparameter settings used for each reported result, justify any differences, and provide a validation procedure (e.g., tune on a subset, report variance across runs or splits).","section":"Sec. IV-E, Sec. V, Fig. 5"},{"comment":"The abstract claims the method 'compares favorably to other state-of-the-art methods,' but the only real-world comparison reported in Sec. V gives FruitNeRF++ an F1-score of 0.765 on the FUJI dataset, whereas Gené-Mola et al. [2] achieve 0.881. The numerical result does not support the phrasing 'compares favorably.' The authors attribute the gap to 'noisy poses and manual registration of both sides,' but no ablation or quantitative analysis is provided to substantiate this explanation. Please either temper the claim to a direct comparison of the reported numbers, or provide supporting experiments that isolate the effect of pose noise (e.g., using refined poses or a subset of images with more reliable registration).","section":"Sec. V, Abstract"},{"comment":"The GT-mask experiments in Table II are a partially circular sanity check: the instance field is trained using ground-truth instance masks (Sec. IV-C.4) and then the counting result is evaluated against the same ground-truth fruit centers. Since the model has been given the correct instance identities during training, the clustering step is essentially verifying that the learned embeddings separate under the same labels used for supervision. These experiments are therefore useful for debugging the clustering and the field, but they cannot be taken as evidence of generalization. The paper should clearly frame the GT-mask results as a sanity check and base claims of practical performance on the SAM/Detic mask experiments, which are the ones that reflect the intended use case.","section":"Table II, Sec. IV-C"}],"minor_comments":[{"comment":"The word 'constrastive' in the phrase 'a constrastive objective' is a typo and should be 'contrastive.'","section":"Sec. II-B"},{"comment":"The phrase 'prevent both gradients to propagate trough the density field' contains a typo: 'trough' should be 'through.'","section":"Sec. IV-C.4"},{"comment":"The sentence 'The results of our counting algorithm compared to Gené-Mola et al. [2] can be attributed to the noisy poses and manual registration of both sides' is grammatically awkward and the assertion is not backed by evidence; please rephrase and provide support or remove the attribution.","section":"Sec. V"},{"comment":"The phrase 'Messy Rooms dataset dataset' has a duplicated word 'dataset.'","section":"Sec. VII"},{"comment":"The caption states 'The best performing segmentation model is highlighted in green,' but the table as rendered does not appear to contain any green highlighting, and the cell 'see caption' is unclear; please ensure the table matches the caption or revise the caption.","section":"Table I"},{"comment":"The experiment on the Messy Rooms dataset is only mentioned in the conclusion with the result 'we detected 99 objects,' with no description of the setup, the number of objects, the metric, or the parameters used; please move this to the evaluation section or remove it.","section":"Sec. VII"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The paper describes a promising method, but the authors overstate the 'fully fruit-agnostic' claim and the abstract's 'compares favorably' is not supported by their own reported F1-scores. The hyperparameter inconsistency between Sec. IV-E and Sec. V is a concrete factual error that must be corrected, and the lack of a validation protocol and error bars is a reproducibility concern. I did not find grounds for rejection because the core contrastive instance-field idea is plausible and the manuscript includes a detailed evaluation on multiple fruit types plus a real-world dataset. With careful revision of the claims, clarification of the hyperparameters, and a proper validation procedure, the paper could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is sound and the paper is honest about limits, but the 'fully fruit-agnostic' claim is overstated and the real-world comparison is less favorable than the abstract suggests.\n\nWhat's genuinely new: replacing FruitNeRF's fruit-specific template matching with a contrastively trained neural instance field, using a prototype-based InfoNCE loss and hard-negative pixel sampling to handle small objects. That is a real change and a sensible direction. The paper is clearly written, the ablations over embedding dimension, temperature, and clustering weight are useful, and the limitations section is candid about mask splitting, training time, and pose sensitivity.\n\nThe soft spots are mostly about claim strength. The abstract says the method 'compares favorably' to other methods, but on the real FUJI dataset FruitNeRF++ gets F1=0.765 versus Gené-Mola's 0.881. The text later admits noisy poses, but the abstract overstates. The 'fully fruit-agnostic' conclusion is also stronger than the evidence: for plum and mango the authors added 'apple' to the Grounded-SAM prompt to get acceptable masks, and with SAM masks recall for those fruits drops to 0.341 and 0.392, while apple/lemon/pear/peach stay near 0.95. That pattern tracks segmentation IoU, not any fruit-independent property of the NeRF or clustering. The neural field and clustering may indeed be shape-agnostic, but the end-to-end pipeline is not demonstrated to be fruit-agnostic. That should be explicitly scoped.\n\nThere are also evaluation issues: the hyperparameters for the FUJI run appear selected after seeing results, with one run and no error bars; the GT-mask experiments are a somewhat circular sanity check; and the Messy Rooms result is one sentence with no details. Code is promised only upon acceptance, so reproducibility is currently limited. These are all fixable with a held-out validation set, multiple runs, and a toned-down conclusion.\n\nThe paper is aimed at researchers working on NeRF-based counting or instance fields. It is a modest but useful advance and deserves serious refereeing: the problem is real, the core idea is plausible, and the evaluation weaknesses are addressable. I would send it to review, but flag the claim overreach and the evaluation as major-revision items.","headline":"A solid extension of FruitNeRF that overstates its fruit-agnostic claim; the core idea is sound, but the evaluation doesn't support end-to-end generalization.","tokens_in":13865,"tokens_out":3293,"would_cite":true,"duration_ms":34781,"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":"FruitNeRF++ claims that fruit counting can be done without fruit-specific templates by learning per-fruit instance embeddings in a neural radiance field.","keywords":["fruit counting","neural radiance fields","contrastive learning","instance segmentation","point cloud clustering","shape-agnostic counting","agricultural computer vision","HDBSCAN"],"falsifier":"On a synthetic scene with ground-truth instance IDs, render leaves that partially occlude one fruit and feed the resulting automatic masks to the pipeline; if the occluded fruit is counted as two while all unobstructed fruits are counted correctly, the load-bearing assumption of mask consistency is falsified for that setting.","tokens_in":12806,"feed_emoji":"🍎","tokens_out":6338,"duration_ms":59182,"temperature":0.7,"pith_summary":"FruitNeRF++ aims to make orchard fruit counting independent of fruit type. Its central claim is that a neural radiance field can learn a per-fruit identity field, trained by contrastive learning on 2D instance masks predicted by vision foundation models, so that counting reduces to clustering a point cloud whose points carry both fruit semantics and instance embeddings. No fruit-specific shape templates are needed, which addresses the main limitation of the prior FruitNeRF method. The paper shows the approach on six synthetic fruit types and a real apple dataset, and argues the same machinery applies to arbitrary objects.","feed_headline":"Counting any fruit type with one NeRF-based pipeline","feed_subtitle":"Instance masks give each fruit a 3D identity, so one clustering step counts apples, plums, mangoes.","key_machinery":"The central object is the neural instance field, a function $F_i$ mapping a 3D point to a $D$-dimensional identity embedding, rendered volumetrically like color and semantics. It is trained with a contrastive loss in which each sampled pixel is attracted to the mean prototype of its own fruit mask and repelled from all other fruit prototypes, with a pixel sampler that mines nearby fruits as hard negatives. A cascaded training scheme first fits density and RGB, then adds the semantic fruit field, then freezes both and trains only the instance field. The final counting step partitions the fruit point cloud into spatial chunks with $k$-Means and clusters each chunk with HDBSCAN using distance $d = \\lambda_c d_c + \\lambda_e d_e$ with $\\lambda_c = 1$ and $\\lambda_e = 5$ in the real-world setting.","core_discovery":"The paper's central discovery is that instance masks, even when they are inconsistent across views, can be fused into a 3D-consistent instance field by contrastive learning: the same fruit's pixels are attracted to a shared prototype embedding while different fruits are repelled. Rendering this field volumetrically and sampling it yields a fruit point cloud with D-dimensional identity features, and clustering with HDBSCAN under a cosine-plus-Euclidean distance gives the count. The authors report that this removes the need for per-fruit templates and that the pipeline, with identical parameters, counts apples, plums, lemons, pears, peaches, and mangoes, plus 99 of 100 common household objects in a separate scene. On the real FUJI apple dataset the F1-score is 0.765 with automatically predicted masks.","pith_inferences":["Because the failure mode is mask splitting, a direct extension is to add a post-hoc merging step: cluster instance embeddings first and then split clusters whose embedding shows a bimodal structure, which could recover fruits split by occluders.","The contrastive prototype loss could be combined with temporal or geometric priors (e.g., fruit centers from the density field) to regularize embeddings for clustered fruits like bananas or berries, which the paper lists as an open challenge.","If the method transfers to arbitrary objects as the Messy Rooms result suggests, it could be used to count other small, partially occluded clusters in the wild, such as flowers, berries on bushes, or manufactured parts in bins, with only text-prompted masks.","The reported dependence on the Euclidean weight $\\lambda_e$ hints that the instance embeddings alone carry less identity signal than the geometry in sparse scenes; on denser or more occluded scenes, higher embedding dimensions and temperature tuning may become critical."],"forward_implications":["The same pipeline, with no fruit-specific parameters, counts apples, plums, lemons, pears, peaches, and mangoes in synthetic scenes, and the authors state the method is agnostic to fruit type in both the neural field and the clustering.","Users no longer need to construct or tune fruit-specific templates or cascaded clustering hyper-parameters, making the method easier to control than FruitNeRF.","Because only RGB images and automatic instance masks are required, the approach extends to arbitrary object types; the authors demonstrate this by detecting 99 of 100 objects in a Messy Rooms household scene.","The counting result depends on the quality of the mask predictions: average F1 over synthetic fruits is 0.925 with ground-truth masks, 0.832 with Grounded-SAM masks, and 0.776 with Detic masks.","The method's speed makes it unsuitable for real-time use: training the FUJI scene takes about 8 hours on an A5000 GPU, with the instance field converging slowly."],"supporting_citations":[{"why":"Supplies the baseline FruitNeRF framework and the synthetic multi-fruit dataset that FruitNeRF++ extends with instance masks.","marker":"[6]"},{"why":"Provides the contrastive learning approach for lifting view-inconsistent 2D instance masks into 3D instance embeddings.","marker":"[22]"},{"why":"Defines the neural radiance field representation and volumetric rendering equations that all fields in FruitNeRF++ build on.","marker":"[23]"},{"why":"Generates the predicted instance masks used to train the instance field in the main experiments.","marker":"[37]"},{"why":"Segment Anything, combined with Grounding DINO, produces the box-prompted masks used for fruit instances.","marker":"[39]"},{"why":"Detic provides an alternative mask and class detector without fine-tuning, used to produce instance masks in the comparison.","marker":"[41]"},{"why":"HDBSCAN is the clustering algorithm applied to the instance-embedded fruit point cloud in the final counting step.","marker":"[48]"},{"why":"Serves as the real-world apple-counting baseline whose F1-score on the FUJI dataset is compared against FruitNeRF++.","marker":"[2]"}],"fun_headline_variants":["One NeRF pipeline counts apples, plums, mangoes, and more","Contrastive learning merges 2D masks into 3D fruit counts","Fruit-agnostic counting: same parameters for six fruit types","Instance embeddings make fruit counting fruit-type-agnostic","From apples to mangoes: single NeRF counting pipeline"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The instance masks from SAM/Detic must assign a consistent identity to each whole fruit across all views; when a leaf or branch splits a fruit into two masks, the contrastive loss pulls the two halves' embeddings apart and the count doubles.","fun_headline_variants_meta":{"raw":{"variants":["One NeRF pipeline counts apples, plums, mangoes, and more","Contrastive learning merges 2D masks into 3D fruit counts","Fruit-agnostic counting: same parameters for six fruit types","Instance embeddings make fruit counting fruit-type-agnostic","From apples to mangoes: single NeRF counting pipeline"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000916,"raw_usage":{"total_tokens":3919,"prompt_tokens":920,"completion_tokens":2999,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":2908}},"tokens_in":536,"tokens_out":2999,"duration_ms":22097,"temperature":1.0,"reasoning_tokens":2908,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:05:37.411751+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a synthetic scene with ground-truth instance IDs, render leaves that partially occlude one fruit and feed the resulting automatic masks to the pipeline; if the occluded fruit is counted as two while all unobstructed fruits are counted correctly, the load-bearing assumption of mask consistency is falsified for that setting.","supporting_citations":[{"cited_title":"FruitNeRF: A Unified Neural Radiance Field based Fruit Counting Framework,","cited_arxiv_id":null,"evidence_quote":"Supplies the baseline FruitNeRF framework and the synthetic multi-fruit dataset that FruitNeRF++ extends with instance masks."},{"cited_title":"Contrastive Lift: 3D Object Instance Segmentation by Slow-Fast Contrastive Fusion,","cited_arxiv_id":null,"evidence_quote":"Provides the contrastive learning approach for lifting view-inconsistent 2D instance masks into 3D instance embeddings."},{"cited_title":"NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis,","cited_arxiv_id":null,"evidence_quote":"Defines the neural radiance field representation and volumetric rendering equations that all fields in FruitNeRF++ build on."},{"cited_title":"Grounded SAM: Assembling Open-World Models for Diverse Visual Tasks,","cited_arxiv_id":null,"evidence_quote":"Generates the predicted instance masks used to train the instance field in the main experiments."},{"cited_title":"Segment Anything,","cited_arxiv_id":null,"evidence_quote":"Segment Anything, combined with Grounding DINO, produces the box-prompted masks used for fruit instances."},{"cited_title":"Detecting Twenty-thousand Classes using Image-level Supervision,","cited_arxiv_id":null,"evidence_quote":"Detic provides an alternative mask and class detector without fine-tuning, used to produce instance masks in the comparison."},{"cited_title":"Density-Based Clustering Based on Hierarchical Density Estimates,","cited_arxiv_id":null,"evidence_quote":"HDBSCAN is the clustering algorithm applied to the instance-embedded fruit point cloud in the final counting step."},{"cited_title":"Fruit detection and 3D location using instance segmentation neural networks and structure-from-motion photogrammetry,","cited_arxiv_id":null,"evidence_quote":"Serves as the real-world apple-counting baseline whose F1-score on the FUJI dataset is compared against FruitNeRF++."}],"review_version":1}