IchthyoNoma: Nomenclature and Context Sensitivity of Zero-Shot Biological Vision--Language Models for Bangladeshi Freshwater Fish Recognition
Nazim-E-Alam, Tarek Rahman, Md Kishor Morol
cs.CV · cs.CL
Quick verdict
Solid zero-shot audit of CLIP/BioCLIP/BioCLIP2 on 7 Bangladeshi fish classes; nomenclature, Bengali and context effects well controlled, scope narrow.
Summary
The paper audits zero-shot recognition by frozen vision-language models on seven freshwater-fish categories shared by two Bangladeshi datasets, BFF-15 (2,656 images) and SylFishBD (7,665 images, with segmentation masks), totaling 10,321 images. Using a four-template prompt ensemble, it compares generic CLIP ViT-B/32, biology-specialized BioCLIP and BioCLIP2, and multilingual Jina CLIP v2 across Romanized, English-common, scientific, and Bengali-script labels. BioCLIP2 is best (72.36% on BFF-15 with English names; 68.91% on SylFishBD with scientific names) versus 25.15%/14.40% for CLIP. Bengali prompts leave BioCLIP2 at chance-level balanced accuracy (14.22-14.29%), while Jina partially recovers Bengali discrimination (21.89%/16.36%) but collapses with bare Bengali names; Jina's English accuracy (29.41%/23.78%) remains far below BioCLIP2. Nomenclature matters: synonym substitutions shift accuracy by up to 8.58 points. Paired context interventions on SylFishBD show weak blur is nonsignificant, medium/strong blur and gray/mean-color masking cost 2.2-3.8 points, white masking costs 8.39, background-only images retain 17.90% accuracy, and cross-species swaps reduce accuracy to 56.76% with low donor-follow (8.88%). Effects are strongly species-dependent. Limitations include only seven classes, a single generic baseline confounded with scale/training recipe, a 512-D truncated Jina representation, unknown pre-training exposure, and Bengali findings tied to the specific names/templates tested. Code and data are on GitHub.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
4 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- BioCLIP2 substantially outperforms generic CLIP on zero-shot recognition of seven Bangladeshi freshwater fish categories (72.36% vs 25.15% on BFF-15 with English prompts; 68.91% vs 14.40% on SylFishBD with scientific prompts).
- Bengali prompts leave BioCLIP2 at chance-level balanced accuracy (14.22-14.29%), while Jina CLIP v2 partially recovers Bengali discrimination (up to 21.89%), showing that multilingual alignment and biological specialization are distinct dimensions.
- Zero-shot accuracy depends on the exact nomenclature and prompt string, with scientific-synonym substitutions shifting accuracy by up to 8.58 points and single-template variation spanning over 20 points.
- Context sensitivity is intervention- and species-dependent: weak blur is nonsignificant, white masking costs 8.39 points, background-only images retain 17.90% accuracy, and donor-follow after cross-species swaps (8.88%) falls below the permutation null (15.10%).
Strengths
Text interface isolated as an explicit variable with a clean within-model design
Section 3.2 caches image embeddings so nomenclature/prompt experiments change only the text-side classifier; Section 4.2 shows BioCLIP2 spans 56.06-77.07% on BFF-15 across four individual scientific templates, and synonym swaps (Gibelion catla +5.14 on SylFishBD; Labeo catla -8.58 on BFF-15) demonstrate the classifier is instantiated by the exact string.
Multilingual control decouples alignment from biological specialization
Section 4.3: Jina CLIP v2 partially recovers Bengali balanced accuracy (21.89% BFF-15, 16.36% SylFishBD) versus BioCLIP2's near-chance 14.22-14.29%, while Jina's English accuracy (29.41%/23.78%) stays far below BioCLIP2; the bare-Bengali-name control shows prompt framing interacts with multilingual ability.
Statistically disciplined intervention ladder with falsification controls
Honest attribution and duplicate handling
Weaknesses
Seven-class ceiling compresses the effect space
Specialization claim rests on one generic baseline
Multilingual control is evaluated handicapped
Bengali failure is under-characterized
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is an audit, not a method paper: four frozen VLMs (CLIP ViT-B/32, BioCLIP, BioCLIP2, Jina CLIP v2) are held fixed while the textual interface (Romanized/English-common/scientific/Bengali labels, four templates, two synonym substitutions) and the visual context (graded blur, neutral masks, crop, background-only, cross-species swaps) are varied over 10,321 images of seven freshwater fish categories shared by BFF-15 (2,656) and SylFishBD (7,665, with SAM-derived masks). Central claims: (1) biology-specialized pretraining dominates generic CLIP (72.36%/68.91% vs 25.15%/14.40% under each dataset's best label family); (2) Bengali prompts leave BioCLIP2 at chance (14.22-14.29% balanced accuracy) while Jina partially recovers Bengali (21.89%/16.36%), decoupling multilingual alignment from biological specialization; (3) exact nomenclature and template are part of the classifier; (4) context robustness is intervention- and species-dependent. Claims are consistently scoped: the paper repeatedly hedges attribution (scale/recipe confound in 4.1; name/template-specific Bengali findings in 4.5).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the IchthyoNoma paper about?
It is an evaluation study asking how much zero-shot fish-recognition scores from vision-language models like CLIP and BioCLIP depend on the class name (Romanized, English common, scientific, or Bengali), the exact prompt wording, and the image background rather than on visual species knowledge alone, using 10,321 images of seven Bangladeshi freshwater fish categories from two public datasets.
Which model performed best for zero-shot fish recognition in this study?
BioCLIP2: 72.36% accuracy on BFF-15 with English common names and 68.91% on SylFishBD with scientific names, versus 25.15% and 14.40% for generic CLIP ViT-B/32. The multilingual Jina CLIP v2 recovered some Bengali discrimination but reached only 29.41%/23.78% with English names, far below BioCLIP2.
Do Bengali prompts work with BioCLIP2 for fish recognition?
No. Bengali-script prompts gave near-chance balanced accuracy (14.22-14.29% across the two datasets, with chance at 14.29% for seven classes). Jina CLIP v2 partially recovered Bengali performance (21.89% and 16.36% balanced accuracy), but with bare Bengali class names even Jina fell back to chance.
Does blurring or masking the background hurt zero-shot fish recognition accuracy?
It depends on the intervention: weak Gaussian blur had no significant effect; medium/strong blur and gray or mean-color masking cost roughly 2-4 points; white masking cost 8.39 points, largely a distribution-shift artifact. Background-only images retained 17.90% accuracy, so the fish dominates the decision but backgrounds carry residual, species-dependent signal.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-3 appear only as captions in the extracted text; per-class values cited here (e.g., Katla +8.91 under strong blur) come from the body prose, not from reading the figures.
- The state of the GitHub repository (code/data completeness) could not be verified from the text; availability is reported as stated by the authors.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.