{"id":"1d75bc40-fa61-4830-a14a-71004aaa1890","arxiv_id":"2505.22850","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A contrastive training loss that groups the image regions most similar to a referring expression, using the true object count, improves counting accuracy by over 22% on REC-8K.","lead":"This paper introduces C-REX, a contrastive learning method for counting objects that match a detailed phrase such as 'person walking' or 'red car'. It reports state-of-the-art accuracy on the REC-8K benchmark, cutting error by over 22% compared to the previous best method.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 80% precision for the top-N positive selection (supp. §6) has no supporting methodology; the entire contrastive gain in Table 4 hinges on this unverified number.","rationale":"The central claim is that the contrastive loss L*_sup delivers the 0.84 MAE gain over the GDino improved baseline (Table 4). The positive set for that loss is chosen by top-N similarity to a masked referring-expression text vector. The paper's only stated evidence that this pseudo-labeling is trustworthy is the supplementary sentence 'approximately 80% of selected samples correspond to correct instances'. No method, tolerance, or dataset supports it. This is a missing support in the exact place where the mechanism must be reliable. The concern is load-bearing because if the true precision is far from 80%, the loss might still help (by grouping attribute-similar queries), but the paper's explanation of why C-REX works would be wrong, and the claim that the choice of N is 'theoretically justified' (Section 3.2) would be unsupported. The problem is compounded by the fact that the positive-selection similarity is image-text (Eq. 2), so the paper's 'entirely within the image space' selling point does not actually cover the mechanism that creates the training signal; the contrastive loss itself is in image space, but the pseudo-labels come from an image-text comparison, which is exactly the alignment the authors say they avoid. I do not think this warrants rejection: the reported numbers on REC-8K may be real and reproducible, and the method could be salvageable with a precision audit and, if needed, a different selection rule. But the current manuscript does not support its own central mechanism as written. A conditional acceptance that requires the precision audit (or code release enabling it) is the appropriate outcome. The reader's conditional verdict already points in this direction; my concern reinforces it and makes the specific missing experiment concrete.","tokens_in":14967,"tokens_out":5667,"duration_ms":56159,"concrete_test":"Run a training-time precision audit on REC-8K: for each image-RE pair, decode the top-N selected queries to their predicted points/boxes and label a query as correct if its prediction falls within a small tolerance of a ground-truth point of the target RE (e.g., distance < 10 px or box IoU > 0.5). Aggregate precision@N over the training set. If the number is not close to 80%, the paper's supportive claim is false. As a causal check, retrain C-REX with N uniformly random tokens as positives; if the test MAE remains near 5.06 rather than reverting toward the 5.90 baseline, the top-N selection is not the source of the gain, and the paper's explanation for its own improvement is called into question.","verdict_should_be":"UNCHANGED","load_bearing_attack":"C-REX's main contribution is the modified supervised contrastive loss L*_sup (§3.2, Eq. 4), which pulls together the N image queries with highest similarity to the masked referring expression (Eqs. 2–3) and pushes away all others. Table 4 shows this loss is what turns the baseline into C-REX (test MAE 5.90 → 5.06). The only stated justification for trusting this pseudo-label selection is in supp. §6: 'approximately 80% of selected samples correspond to correct instances.' No experiment, metric, tolerance, or dataset is given for that number, so the central training signal is supported by an unverifiable assertion. This matters doubly because the selection operator itself is an image-text similarity (zi·tm), so the paper's claim that C-REX 'operates entirely within the image space' and avoids image-text misalignment is only true for the loss term, not for the mechanism that generates the positives. If actual precision is well below 80%, the loss could be clustering queries that merely share attribute-level similarity with the text, which may still help counting but would not support the stated mechanism or the 'theoretically justified' choice of N.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes C-REX, a contrastive learning framework for referring expression counting (REC). It builds on a Grounding DINO-based baseline that is repurposed from bounding-box prediction to object-centroid prediction, and adds a modified supervised contrastive loss (L*_sup) that treats the N image tokens most similar to the attribute-only referring expression as positives, where N is the ground-truth count. The authors report state-of-the-art results on REC-8K (test MAE 5.06 vs. 6.50 for GroundingREC) and competitive performance on class-agnostic counting benchmarks (FSC-147, CARPK). The main claimed contributions are: (1) identifying centroid detection as the key factor in recent detection-based counting models, (2) the proposed modified supervised contrastive loss, and (3) the versatility of the framework beyond REC.","tokens_in":15240,"tokens_out":7159,"duration_ms":75208,"significance":"If the results hold, the paper makes a useful empirical contribution: it provides a strong detection-based REC baseline and shows that a supervised contrastive loss operating on image tokens, with pseudo-positives selected by text similarity, can substantially improve counting accuracy. The released code and the clear ablations in Tables 4-6 are strengths. The empirical gain over the previous REC-specific method GroundingREC is large (22% relative MAE improvement on the test set). However, several load-bearing pieces of evidence are incomplete: the claimed 80% precision of the pseudo-label selection is unverified, the centroid-detection conclusion is not isolated in a controlled ablation, and the assertion that C-REX operates 'entirely within the image space' is overstated because positive selection uses text-image similarity.","major_comments":[{"comment":"The statement that 'approximately 80% of selected samples correspond to correct instances' is presented without any methodology, experiment, or error tolerance to support it. This number is used to justify the top-N selection strategy and the choice of N as the ground-truth count, so it is load-bearing for the proposed training signal. Please provide the verification procedure and results, or remove the quantitative claim and soften the 'theoretically justified' language. Additionally, the label-noise citations appear as unresolved placeholders '[? ? ?]'.","section":"Supplementary §6"},{"comment":"The claim that C-REX 'operates entirely within the image space' is contradicted by the positive-selection mechanism: in Eq. (2), pseudo-labels are assigned by computing cosine similarity between image tokens z_i and the masked referring-expression text token t_m. The contrastive loss is indeed computed in image space, but the pseudo-labeling still relies on image-text alignment. This undermines the stated advantage over image-text contrastive learning (e.g., avoiding misalignment, stability). Please qualify the claim to say that only the contrastive objective is image-space, and discuss how the quality of the frozen text encoder's alignment affects the selection.","section":"Eq. (2) and §3.4"},{"comment":"The conclusion that 'detecting object centroids instead of bounding boxes is the key common factor' behind the success of detection-based counting models is not supported by the provided ablation. Table 6 compares full GroundingREC, GroundingREC without feature fusion, and the proposed GDino improved baseline, but these variants differ simultaneously in output format (box vs. point), presence of the contrastive module, and feature-fusion components. The text says the CL module was removed to form the baseline, but that intermediate row is not shown. To support the claim, please add an ablation that holds the architecture fixed and varies only the output head (bounding box vs. centroid), and show the incremental effect of removing feature fusion and contrastive learning from GroundingREC.","section":"§4.5.3 / Table 6"},{"comment":"The modified supervised contrastive loss L*_sup is undefined when the positive set I+ contains exactly one element (i.e., N=1): the inner average over p in I+ - {i} is empty and the 1/(|I+|-1) factor divides by zero; for N=0, the outer sum is empty. The paper does not specify how such anchors are handled during training. Please state the implementation detail (e.g., skipping anchors with no other positive, or excluding samples with N<2) and confirm that training is stable for the distribution of counts in REC-8K.","section":"Eq. (4)"}],"minor_comments":[{"comment":"There are several typographical and wording errors: 'choise' in §4.5.2, 'Noteably' in §4.3, 'differencese.g.' in the Introduction, and informal uses of 'sota' throughout. Please proofread the manuscript.","section":"Global"},{"comment":"The unresolved citation placeholders '[? ? ?]' should be replaced with proper references for label noise in contrastive learning.","section":"Supplementary §6"},{"comment":"The F1 score is reported for REC but its formula is not defined anywhere, including in the supplementary where it is only described verbally. Please provide the exact computation, including how true positives are matched to ground-truth points.","section":"§4.2"},{"comment":"The detection thresholds 0.30 (CLS token) and 0.36 (other text tokens) are stated without a sensitivity analysis or a description of how they were chosen. Please clarify whether these were selected on the validation set and how robust the results are to them.","section":"§4.1"},{"comment":"The description of the GDino improved baseline is sparse: it says bounding-box prediction is replaced by point-center prediction, but it is not specified which architecture changes are made, how the point head is attached, or which tokens are supervised by the L1 point-regression loss. More detail would aid reproducibility.","section":"§3.1"},{"comment":"The phrase 'theoretically justified' for the choice of N is stronger than what the supplementary actually provides, which is an intuitive argument about decision boundaries. Please align the wording with the level of support given.","section":"§1 and §4.5.2"}],"recommendation":"major_revision","confidential_remarks":"The paper reports strong empirical results and the code is released, which is a positive sign. However, the authors' key explanatory claims—the 80% pseudo-label precision and the centroid-detection conclusion—require more rigorous support. The 'entirely within the image space' characterization is likely to be challenged by reviewers and should be corrected. The paper is within the scope of the journal and appears fixable, but the revisions are substantive rather than cosmetic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is a supervised contrastive loss that runs on image tokens, uses the top-N most text-similar tokens as positives (N from the ground-truth count), and only anchors on positives. The REC-8K numbers are a real jump: test MAE 5.06 vs. 6.50 for GroundingREC, and the ablations in Table 4 show the modified loss is doing the work, not the baseline. The improved GDino baseline also stands on its own. They released code, they retrained CountGD for a fair comparison, and they show failure cases. That is honest, reproducible work and deserves credit.\n\nThe soft spots are real but not fatal. First, the supplementary's \"approximately 80% of selected samples correspond to correct instances\" has no methodology behind it. No experiment, no metric, no tolerance. That matters because the contrastive loss's pseudo-labels come from this selection. It doesn't make the result circular—N is a training signal, and test counts come from detector outputs—but it does undercut the claim that N is \"theoretically justified.\" A referee should ask them to actually measure selection precision. Second, the \"centroid detection is the key factor\" insight is not cleanly isolated: GDino improved differs from GroundingREC in several ways at once, not just the readout head. Third, no variance or seed information anywhere, so the gap could be part noise. Fourth, the \"operates entirely within the image space\" slogan is only true for the loss; the positive-selection mechanism itself uses image-text similarity (zi·tm), so the text embedding is still doing work inside the training loop. That's a framing fix, not a flaw in the method.\n\nThe stress-test concern about the 80% number is legitimate, and I want it named, but it does not sink the paper. The empirical gain exists, the loss is well specified, and the missing measurement is easy to add. The reader's conditional verdict matches my own: plausible, needs code and repeated runs, but worth engaging.\n\nFor a reviewer: this should go to peer review. The core contribution is new, the experiments are informative, and the weaknesses are patchable rather than structural. I would cite it if I worked in counting, and I'd bring it to reading group as a useful example of image-space contrastive learning applied to a fine-grained task.","headline":"A solid REC paper with a genuine empirical gain, one unverified supporting number in the supplementary, and an overreaching 'key factor' claim; the core method is worth refereeing and likely worth citing.","tokens_in":15741,"tokens_out":1681,"would_cite":true,"duration_ms":18686,"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":"A contrastive loss that stays in image space cuts referring-expression counting error by 22%.","keywords":["referring expression counting","contrastive learning","object counting","class-agnostic counting","open-set detection","centroid detection","supervised contrastive loss","REC-8K"],"falsifier":"Retrain C-REX exactly as described but replace the top-$N$ similarity selection with $N$ randomly chosen image tokens, keeping all other settings fixed; if REC-8K test MAE stays near 5.06 rather than degrading toward the 5.90 baseline, the similarity-based selection of positives is not what drives the gain. A supporting measurement is the actual precision of top-$N$ selection on held-out images: if it is far below the stated 80%, the loss is being trained on systematically wrong positives.","tokens_in":14813,"feed_emoji":"🔢","tokens_out":6797,"duration_ms":64898,"temperature":0.7,"pith_summary":"The paper proposes C-REX, a contrastive learning method for Referring Expression Counting (REC), the task of counting only the objects in an image that match a fine-grained description such as \"person walking\" or \"car driving to the right.\" Its central claim is that the contrastive signal should live entirely in image space: the model selects the $N$ image tokens most similar to the referring expression's text token, where $N$ is the ground-truth count, treats those as positives, and treats every other image token as a negative, then pulls positives together while pushing negatives away. Because the negative pool is all image tokens rather than a handful of text expressions, the loss is both larger and more stable than image-text contrastive learning. On the REC-8K benchmark the authors report state-of-the-art results, improving MAE by over 22% and RMSE by over 10% over the previous best. A reader should care because distinguishing visually similar objects that belong to different referring expressions is the bottleneck for the next generation of counting models, and C-REX shows that a supervision signal as cheap as the count itself can break that bottleneck.","feed_headline":"Image-space contrastive loss cuts counting error by 22%","feed_subtitle":"C-REX picks the ground-truth count of most-similar image tokens as positives and beats prior referring-expression counters.","key_machinery":"The load-bearing object is the modified supervised contrastive loss $L^*_{sup}$ (Eq. 4), which uses only positive-class image tokens as anchors: each anchor is pulled toward its selected positive peers and pushed away from all other image tokens. The positive set is produced by a cosine-similarity ranking between image tokens and the masked referring-expression token, with the ground-truth count $N$ fixing how many top tokens become positives (Eqs. 2-3). The supporting machinery is the centroid-prediction baseline: Grounding DINO is fine-tuned with an L1 point-center loss and a cross-entropy classification loss instead of bounding-box regression, which the paper argues is the key component behind state-of-the-art detection-based counting.","core_discovery":"On its own terms, the paper establishes that the failure of earlier REC models to tell apart visually similar objects of the same class is largely a representation-learning problem, and that a modified supervised contrastive objective fixes it without any extra annotation. The method, C-REX, is built on a detection baseline that repurposes the open-set detector Grounding DINO from bounding-box prediction to object-centroid prediction, which the authors identify as the common success factor behind recent detection-based counters. On top of that baseline, the contrastive module computes a masked referring-expression token, ranks image tokens by cosine similarity, and labels the top $N$ as positives using the ground-truth count as $N$. The loss uses only positive tokens as anchors, so diverse negatives are not erroneously pulled together. Experiments on REC-8K show test MAE 5.06 versus 6.50 for GroundingREC and test RMSE 17.53 versus 19.79, with consistent gains across count ranges and attribute categories; a class-agnostic counting variant also matches or beats prior text-based methods on FSC-147.","pith_inferences":["Beyond the paper: the reported \"about 80%\" precision of top-$N$ selection implies the loss tolerates roughly one in five noisy positives; a useful next experiment is to measure how this precision varies with object count, since the 100+ bin shows the smallest RMSE gain.","Beyond the paper: the claim that image-space contrast is inherently more stable than image-text contrast still awaits a matched comparison with equal negative counts, because the paper removes GroundingREC's feature fusion but does not directly ablate its image-text contrastive loss in the same pipeline.","Beyond the paper: the top-$N$ contrastive recipe should transfer to referring expression segmentation, where positive tokens could be selected from image features overlapping the target mask; this follows from the paper's generality claim but is not tested.","Beyond the paper: if top-$N$ precision stays high for rare attribute categories such as gender and orientation, C-REX should generalize to unseen referring expressions at inference time, since no new positive labels would be required."],"forward_implications":["On REC-8K, C-REX sets a new state of the art with test MAE 5.06 and RMSE 17.53, beating GroundingREC by roughly 22% and 11% respectively; future REC systems should treat image-space contrastive positives as a strong default.","The only extra supervision beyond the existing count labels is the referring expression itself, so the method scales to datasets that have text prompts and counts but no bounding-box or mask annotations.","The centroid-detection baseline alone, without contrastive learning, already beats GroundingREC on MAE, so point-center prediction is a cheap upgrade for detection-based counting architectures.","The positive-only anchor modification matters: replacing $L^*_{sup}$ with the unmodified supervised contrastive loss raises test MAE from 5.06 to 5.91, identifying anchor choice as a key ingredient.","The same recipe transfers to class-agnostic counting by using class similarity instead of referring-expression similarity, and the resulting text-based results on FSC-147 are competitive with prior text-based methods."],"supporting_citations":[{"why":"Introduces the REC-8K dataset and task, the GroundingREC baseline, and the image-text contrastive approach that C-REX replaces.","marker":"[11]"},{"why":"Provides the Grounding DINO open-set detector whose frozen backbone, feature enhancer, and decoder tokens C-REX fine-tunes and contrasts.","marker":"[28]"},{"why":"Supplies the supervised contrastive loss that C-REX modifies into a positive-only-anchor objective.","marker":"[24]"},{"why":"CountGD is the detection-based state-of-the-art method whose box-to-centroid design the authors analyze to motivate their baseline.","marker":"[3]"},{"why":"Supports the paper's claim that image-space contrastive learning is more data-efficient than image-text contrastive learning.","marker":"[34]"},{"why":"FSC-147 is the class-agnostic counting benchmark used to evaluate the C-REX adaptation.","marker":"[42]"},{"why":"DINO's self-supervised image representations underpin the image-space contrastive efficiency argument and the detection backbone lineage.","marker":"[57]"}],"fun_headline_variants":["C-REX: Image-space contrastive learning cuts counting error 22%","Centroid-based contrastive model improves referring expression counting","Image-space contrastive learning SOTA for referring expression counting","C-REX beats prior REC models with 22% lower MAE and 10% lower RMSE","Contrastive learning in image space improves referring expression counting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole contrastive signal rests on the similarity ranking being correct: the paper assumes the $N$ image tokens most similar to the referring expression are the objects it names, reporting only about 80% correct selections and giving no measured rule for when that ranking fails.","fun_headline_variants_meta":{"raw":{"variants":["C-REX: Image-space contrastive learning cuts counting error 22%","Centroid-based contrastive model improves referring expression counting","Image-space contrastive learning SOTA for referring expression counting","C-REX beats prior REC models with 22% lower MAE and 10% lower RMSE","Contrastive learning in image space improves referring expression counting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00084,"raw_usage":{"total_tokens":3714,"prompt_tokens":1052,"completion_tokens":2662,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":668,"completion_tokens_details":{"reasoning_tokens":2566}},"tokens_in":668,"tokens_out":2662,"duration_ms":21957,"temperature":1.0,"reasoning_tokens":2566,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:59:01.897819+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain C-REX exactly as described but replace the top-$N$ similarity selection with $N$ randomly chosen image tokens, keeping all other settings fixed; if REC-8K test MAE stays near 5.06 rather than degrading toward the 5.90 baseline, the similarity-based selection of positives is not what drives the gain. A supporting measurement is the actual precision of top-$N$ selection on held-out images: if it is far below the stated 80%, the loss is being trained on systematically wrong positives.","supporting_citations":[{"cited_title":"Referring ex- pression counting","cited_arxiv_id":null,"evidence_quote":"Introduces the REC-8K dataset and task, the GroundingREC baseline, and the image-text contrastive approach that C-REX replaces."},{"cited_title":"Grounding dino: Marrying dino with grounded pre-training for open-set object detection.European Confer- ence on Computer Vision, 2024","cited_arxiv_id":null,"evidence_quote":"Provides the Grounding DINO open-set detector whose frozen backbone, feature enhancer, and decoder tokens C-REX fine-tunes and contrasts."},{"cited_title":"Supervised contrastive learning.Advances in neural information processing systems, 33:18661–18673,","cited_arxiv_id":null,"evidence_quote":"Supplies the supervised contrastive loss that C-REX modifies into a positive-only-anchor objective."},{"cited_title":"Amini-Naieni, T","cited_arxiv_id":null,"evidence_quote":"CountGD is the detection-based state-of-the-art method whose box-to-centroid design the authors analyze to motivate their baseline."},{"cited_title":"Dinov2: Learning robust visual features without supervision","cited_arxiv_id":null,"evidence_quote":"Supports the paper's claim that image-space contrastive learning is more data-efficient than image-text contrastive learning."},{"cited_title":"Learning to count everything","cited_arxiv_id":null,"evidence_quote":"FSC-147 is the class-agnostic counting benchmark used to evaluate the C-REX adaptation."},{"cited_title":"DINO: DETR with improved denoising anchor boxes for end-to-end object de- tection","cited_arxiv_id":null,"evidence_quote":"DINO's self-supervised image representations underpin the image-space contrastive efficiency argument and the detection backbone lineage."}],"review_version":1}