{"id":"4ec779ac-4fb5-41ce-a1bd-2cc42ab088e2","arxiv_id":"2505.15398","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"RichCount improves zero-shot object counting by enriching text prompts with MLLM-generated descriptions and aligning them to CLIP visual features, achieving state-of-the-art mean absolute error on three counting benchmarks.","lead":"RichCount trains a CLIP-based counter in two stages, enriching category labels with ChatGPT-4 descriptions and aligning text and image features with an adapter and feed-forward network. On FSC-147, CARPK, and ShanghaiTech it reports lower counting errors than prior zero-shot text-based counters.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline 15.78 MAE depends on ChatGPT-4 descriptions generated from the test image itself; the paper's label-only variant is still competitive, so the concern is about protocol and claim strength, not internal consistency.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing concern: the best results use image-conditioned MLLM descriptions at test time, which departs from the usual zero-shot protocol where only a category label is available. I agree with the reader's conditional verdict rather than escalating to rejection, because the paper's Text-only variant (16.37) still outperforms the reproduced CLIP-Count baseline (17.19) and is competitive with prior zero-shot methods. The concern is therefore about the strength and interpretation of the headline claim, not about the soundness of the two-stage training pipeline. A controlled comparison of image-conditioned versus image-free descriptions is the decisive check: it isolates whether the description advantage comes from richer semantics or from test-image information such as color and location. If the advantage persists without image access, the central claim is largely vindicated; if not, the paper should be reported as a label-only method with a smaller gain and the SOTA claim should be softened. The lack of released code, generated prompts, and error bars makes this verification necessary rather than optional. For these reasons, the appropriate verdict remains CONDITIONAL, matching the reader's judgment, and no verdict adjustment is needed.","tokens_in":14276,"tokens_out":5471,"duration_ms":53526,"concrete_test":"Freeze the RichCount checkpoint and evaluate on the FSC-147 test split under three text conditions: (1) class labels only; (2) ChatGPT-4 descriptions generated from the full test image as in the paper; (3) ChatGPT-4 descriptions generated from the category name only, without the test image, or from a different image of the same category. Fix the API temperature and run each condition at least 3 times. If condition (2) does not beat condition (3) by more than the run-to-run uncertainty, the 15.78 headline is not evidence of prompt enrichment but of test-image access. If condition (3) ≈ condition (2), the concern is resolved. Report MAE/RMSE and release all generated prompts.","verdict_should_be":"UNCHANGED","load_bearing_attack":"RichCount's headline result (FSC-147 test MAE 15.78, Table 1) is obtained with descriptions T_d = G(I, T_p) generated by ChatGPT-4 from the test image itself (Sec. 3.1, Eq. 4; Table 5). The paper's stated zero-shot condition only requires disjoint train/val/test categories (Sec. 3), so the method is internally consistent as defined. The load-bearing problem is external validity: at test time the textual prompt can encode the image's color, location, and layout, so the model is not counting from a user-supplied label alone. The reported 0.59 MAE gain of Description over Text (15.78 vs 16.37) is exactly the margin that may be attributable to test-image information in the prompt rather than to text enrichment per se. The paper's own Text-only variant still outperforms prior baselines, so this does not refute the method; it weakens the specific 'state-of-the-art 15.78' claim and the claim that rich-prompt generalization is what drives the gain. Additionally, the descriptions come from a non-deterministic external API, no code or generated prompts are released, and no error bars are given, so the exact prompt protocol and its variance cannot be audited.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RichCount, a two-stage framework for zero-shot object counting. In the first stage, a feed-forward network (FFN) and a text adapter are trained with a contrastive loss to align CLIP visual and textual features, using both category labels and MLLM-generated image descriptions as text inputs. In the second stage, the encoders are frozen and an interaction module plus a decoder are trained to produce density maps from diverse textual prompts. Experiments on FSC-147, CARPK, ShanghaiTech, and COUNTBENCH report state-of-the-art zero-shot counting results, with a headline test MAE of 15.78 on FSC-147 obtained with ChatGPT-4 generated descriptions.","tokens_in":14446,"tokens_out":8895,"duration_ms":73727,"significance":"The two-stage alignment recipe (FFN on the visual encoder, adapter on the text encoder, contrastive loss with margin) is a plausible and potentially useful contribution, and the paper's own text-only variant (FSC-147 test MAE 16.37) still beats the reproduced CLIP-Count baseline (17.19), suggesting that the alignment mechanism has some value independent of description augmentation. The paper also provides ablations, cross-dataset transfer experiments, and an analysis of different MLLMs, which are informative. However, the headline result is obtained with test-image-derived descriptions, which compromises the zero-shot claim; the lack of error bars, released prompts, and code limits reproducibility; and the ablation does not cleanly isolate the alignment contribution from the description contribution. These issues are load-bearing for the central claim of state-of-the-art zero-shot counting.","major_comments":[{"comment":"The headline FSC-147 test MAE of 15.78 reported in Table 1 for \"RichCount (Ours) Description\" is obtained with descriptions T_d = G(I, P_t) generated by ChatGPT-4 from the test image I itself (Eq. (4)). Because the description can encode the target objects' color, location, and layout, the textual prompt carries information that is unavailable to the user-supplied category label in a conventional zero-shot setting. The comparison to prior zero-shot methods (CLIP-Count, CounTX, etc.) that use only category names is therefore not a like-for-like zero-shot evaluation, and the paper's claim of \"state-of-the-art performance in zero-shot counting\" (Abstract; Sec. 4.2) is overstated. The paper's own \"RichCount (Ours) Text\" row (test MAE 16.37) is the appropriate zero-shot comparison and still improves on the reproduced CLIP-Count baseline (17.19), but the 0.59 MAE difference between the Description and Text rows is potentially attributable to test-image information rather than to the proposed text-image alignment. Please present the Text-only result as the primary zero-shot claim, or explicitly redefine the protocol to allow image-derived text and compare against methods that use similar image-derived prompts.","section":"Section 3.1, Eq. (4); Table 1"},{"comment":"The experimental protocol is not auditable. The exact prompt template P_t for the MLLM is not specified (only a single example appears in Fig. 2), no generated descriptions are released, no code or seeds are provided, and no error bars or multiple runs are reported for any result in Tables 1-5. Since T_d is the output of a proprietary, non-deterministic model (ChatGPT-4), the reported test MAE of 15.78 is a single sample of a random quantity; the 0.59 MAE advantage over the Text variant may be within run-to-run variation. The authors should release the prompt templates and generated descriptions, specify the API version and sampling parameters (e.g., temperature), and report means and variances over multiple runs for the headline comparisons.","section":"Section 3.1; Table 5"},{"comment":"The ablation study in Table 4 does not separate the contribution of the proposed alignment modules (FFN, adapter, contrastive loss) from the contribution of the image-derived descriptions. The text states that \"the Adapter ... and the contrastive loss were the most influential factors,\" but no condition combines these components with text-only descriptions (e.g., category names) during training and inference. Without such a condition, the observed improvement can be attributed either to the alignment mechanism or to the extra information in the descriptions. Please add an ablation with the full alignment pipeline but without Des (and with text-only inference) to isolate the effect of the proposed modules.","section":"Section 4.3, Table 4"}],"minor_comments":[{"comment":"Table captions and cross-references: Sec. 4.2 refers to \"Tab. 6\" for the CARPK results, which is Table 2; Sec. 4.3 refers to \"Tab. 7\" and \"Tab. 8\" for the component and MLLM ablations, which are Tables 4 and 5; the supplementary text has similar mismatches. Please renumber or fix the cross-references.","section":"Section 4.2-4.3; Table captions"},{"comment":"The prose after Eq. (14) introduces an index x that does not appear in the formula; the sums are indexed by a and (a,b). Please clarify the notation or revise the explanatory sentence.","section":"Equation (14)"},{"comment":"The right panel is labeled \"EACount(Ours)\" but the paper's model is RichCount, and \"B/L\" is not defined. Please correct the label and define the abbreviation.","section":"Figure 5 caption"},{"comment":"The phrase \"Following ClipViT\" uses an undefined term; presumably it means CLIP ViT. Also, the statement that the image and text encoders are \"fixed\" is confusing because the FFN and adapter are trainable; please rephrase.","section":"Section 4.1"},{"comment":"FSC-147 is cited as [7] (the RCC paper by Hobley and Prisacariu), but the dataset was introduced in the FamNet paper [22]. Please correct the dataset citation in Sec. 4.1 and in the supplementary material.","section":"Section 4.1; References"},{"comment":"The abstract's claim of \"the first framework to address these limitations\" is too strong given prior text-description-based counting work such as CounTX; please soften or substantiate the novelty claim. Additionally, in Table 4, the meaning of the \"#\" symbols is not defined: the first row has four \"#\"s and reports baseline-level numbers, while the last row has no \"#\" and reports the best numbers, making it impossible to know which configuration is the full model. Please define the notation.","section":"Abstract; Table 4"}],"recommendation":"major_revision","confidential_remarks":"The paper has a working method and the text-only variant is a defensible zero-shot claim, but the current headline result is overstated because it relies on test-image-derived descriptions. The authors should be asked to reframe the contribution as either zero-shot counting with category labels (using the Text row as the primary result) or as text-guided counting with image-derived descriptions, with the protocol explicitly stated. Releasing the prompt templates and generated descriptions, and reporting variance across MLLM runs, is essential for the claim to be verifiable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the headline number is real but narrower than it looks. The 15.78 FSC-147 test MAE comes from ChatGPT-4 descriptions generated from the test image itself (Sec. 3.1, Eq. 4), which puts test-image content into the prompt — a deviation from the usual zero-shot protocol where the user gives only a category label. The paper never flags this. The good news is the concern is about claim strength, not internal consistency: the Text-only variant still hits 16.37 against the reproduced CLIP-Count's 17.19, so the alignment machinery helps without the descriptions.\n\nWhat's new here is the two-stage training: an FFN on the CLIP visual encoder and an adapter on the text encoder are aligned with a contrastive loss on category, description, and 'object'-replaced prompts; then the frozen encoders feed a counting head trained with consistency losses across those three prompt forms. The 'object' replacement and the cross-prompt consistency are the interesting pieces, and the ablations earn their place. Table 4 plus the supplementary margin and FFN/adapter sweeps are honest and fairly complete, and the gains are consistent across FSC-147, CARPK, ShanghaiTech, and COUNTBENCH.\n\nSoft spots, in order of severity. First, the test-time MLLM: the 0.59 MAE Description-over-Text gap is exactly the margin that could come from image-derived color and location in the prompt rather than from text enrichment per se. The prompt template appears in Fig. 2 and asks for a description of the target in the photo, so the test-time protocol is genuinely different from the training-time one described in Sec. 3. Second, nothing is released — no code, no generated descriptions, and the API is non-deterministic, so with no error bars the exact numbers cannot be audited. Third, 'first framework' in the abstract is overclaimed; every ingredient exists, and the combination, while useful, is a combination. Fourth, the gains are modest: roughly 8–10% relative MAE.\n\nWho it is for: people working on zero-shot counting and CLIP-based alignment. It should go to peer review — the engineering is clean, the ablation story is coherent, and the text-only result shows the core method stands on its own. The referee should require the label-only setting as the primary zero-shot result, an error-bar or multi-run table, and ideally release of the generated descriptions or the prompt template. I would bring it to the reading group to argue about what 'zero-shot' should mean when the test-time prompt is generated from the image.","headline":"Real but narrower than it looks: the 15.78 headline uses ChatGPT-4 descriptions generated from the test image itself, yet the text-only variant still beats baselines — a protocol concern on top of a solid paper that deserves review.","tokens_in":15096,"tokens_out":3873,"would_cite":true,"duration_ms":30756,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"RichCount claims that zero-shot object counting improves when simple category labels are replaced with rich, image-specific descriptions and the text and image encoders are aligned in a dedicated first training stage.","keywords":["zero-shot object counting","text-guided counting","vision-language alignment","CLIP","density map","class-agnostic counting","multimodal language model","prompt enrichment"],"falsifier":"Take the FSC-147 test split, strip the image-specific portions from the descriptions or replace them with generic category descriptions, and re-run RichCount; if the mean absolute error rises to the level of the category-label baseline, the reported gain is caused by information leaked from the test image into the prompt rather than by the alignment modules.","tokens_in":13993,"feed_emoji":"🔢","tokens_out":6567,"duration_ms":53839,"temperature":0.7,"pith_summary":"The paper argues that zero-shot object counting fails with simple category labels because text and image features are misaligned and category names carry too little semantic content. It proposes RichCount, a two-stage training strategy: first enrich category labels into detailed, image-specific descriptions and align those descriptions to visual features with a feed-forward network and an adapter trained under a contrastive margin loss; then freeze those encoders and train a text-conditioned density-map counter. On FSC-147 the method reports a test-set mean absolute error of 15.78, better than the reproduced CLIP-Count baseline at 17.19, and it reports improved transfer to car and crowd datasets. The paper's message is that prompt enrichment alone is not enough; the richer text has to be actively aligned with visual features before it helps counting.","feed_headline":"Replacing labels with rich prompts cuts zero-shot count error","feed_subtitle":"Enriched image descriptions aligned with a CLIP encoder beat text-only baselines and transfer across datasets.","key_machinery":"The central mechanism is a two-stage alignment-and-count pipeline built on CLIP, a pre-trained vision-language encoder. Stage one augments the frozen visual encoder with a feed-forward network and the frozen text encoder with an adapter, and trains them with a contrastive margin loss so that images, category labels, rich descriptions, and negative descriptions separate properly in a shared embedding space. Stage two freezes those aligned encoders and trains an interaction module that treats image features as queries and text features as keys and values, followed by a decoder that regresses density maps; a consistency loss requires predictions from category labels, rich descriptions, and generic 'object' descriptions to agree. This machinery converts a one-step text-to-count problem into a two-step text-alignment-then-count problem.","core_discovery":"On the paper's own terms, RichCount establishes that the bottleneck in zero-shot counting is cross-modal alignment rather than prompt vocabulary. Replacing a category name like 'bread rolls' with a description such as 'golden-brown bread rolls on two metal baking trays on a stove top' only helps if the text encoder and image encoder are first tuned so that such descriptions sit close to the corresponding image regions. The paper trains a feed-forward network on the visual side and an adapter on the text side using a contrastive loss with a margin, then freezes both encoders and trains an interaction module and decoder to output density maps from category labels, detailed descriptions, and generic 'object' descriptions, with a loss that forces all three prediction modes to agree. The reported result is state-of-the-art zero-shot counting on FSC-147, with the best configuration using image-specific descriptions.","pith_inferences":["Because the alignment stage is trained only on image-text similarity, it should carry over to other text-conditioned dense prediction tasks such as referring segmentation or open-vocabulary detection; a test would be to take the aligned encoders and swap the counting head for a segmentation head.","The reported advantage may partly come from descriptions that reference the test image itself, including its colors, layout, and object locations; a cleaner zero-shot evaluation would hold out such image-specific descriptions or generate them only during training.","If a smaller model can generate descriptions of comparable accuracy, the method's dependence on an expensive external multimodal language model could be removed; the paper's own text-only variant is weaker, so the cost-quality trade-off is the decisive open question."],"forward_implications":["Zero-shot counting can move beyond fixed category labels to arbitrary free-text prompts such as attributes, questions, and spatial descriptions during inference.","The same aligned text encoder is shared between the alignment stage and the counting stage, so the counting model does not need to relearn text-vision associations for every new category.","Cross-dataset transfer improves: trained on FSC-147, RichCount reports an MAE of 9.91 on CARPK, below the 13.59 baseline, and lower error than CLIP-based baselines on the ShanghaiTech crowd datasets.","The consistency loss among category, description, and generic-description predictions makes the model less sensitive to the exact wording of the prompt at test time."],"supporting_citations":[{"why":"provides the CLIP-Count text-guided counting baseline that RichCount reproduces and improves on.","marker":"[11]"},{"why":"supplies the multimodal language model used to generate the image-specific descriptions at the core of the method.","marker":"[19]"},{"why":"the marker the paper cites for FSC-147, the main benchmark dataset, and for RCC, the reference-less text baseline.","marker":"[7]"},{"why":"CounTX is the text-description counting baseline that RichCount argues does not fully close the modal gap.","marker":"[1]"},{"why":"VLCount is the text-aware representation method whose alignment strategy RichCount extends.","marker":"[12]"},{"why":"ZSC is a zero-shot text-plus-exemplar counting baseline compared in the experiments.","marker":"[30]"},{"why":"VA-Count is the visual-exemplar zero-shot baseline against which RichCount positions its text-only approach.","marker":"[35]"}],"fun_headline_variants":["Text-vision alignment, not richer prompts, drives zero-shot counting","Rich prompts cut zero-shot count error only after CLIP alignment","Zero-shot counting improved by aligning text and visual features","RichCount: align text encoder to count unseen categories accurately","Why rich prompts alone fail at zero-shot object counting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's best results rely on an external multimodal language model looking at each test image and producing an accurate, image-specific description of the objects to be counted, which goes beyond the usual zero-shot setting where only a category label is available.","fun_headline_variants_meta":{"raw":{"variants":["Text-vision alignment, not richer prompts, drives zero-shot counting","Rich prompts cut zero-shot count error only after CLIP alignment","Zero-shot counting improved by aligning text and visual features","RichCount: align text encoder to count unseen categories accurately","Why rich prompts alone fail at zero-shot object counting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000295,"raw_usage":{"total_tokens":1679,"prompt_tokens":876,"completion_tokens":803,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":492,"completion_tokens_details":{"reasoning_tokens":721}},"tokens_in":492,"tokens_out":803,"duration_ms":7132,"temperature":1.0,"reasoning_tokens":721,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:17:50.281700+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the FSC-147 test split, strip the image-specific portions from the descriptions or replace them with generic category descriptions, and re-run RichCount; if the mean absolute error rises to the level of the category-label baseline, the reported gain is caused by information leaked from the test image into the prompt rather than by the alignment modules.","supporting_citations":[{"cited_title":"Clip-count: Towards text-guided zero-shot object counting","cited_arxiv_id":null,"evidence_quote":"provides the CLIP-Count text-guided counting baseline that RichCount reproduces and improves on."},{"cited_title":"Vlcounter: Text-aware visual representation for zero- shot object counting","cited_arxiv_id":null,"evidence_quote":"VLCount is the text-aware representation method whose alignment strategy RichCount extends."},{"cited_title":"Zero-shot object counting","cited_arxiv_id":null,"evidence_quote":"ZSC is a zero-shot text-plus-exemplar counting baseline compared in the experiments."},{"cited_title":"Zero-shot object counting with good exemplars","cited_arxiv_id":null,"evidence_quote":"VA-Count is the visual-exemplar zero-shot baseline against which RichCount positions its text-only approach."}],"review_version":1}