Tpc 268 Eval

tpc-268-eval

Plant Taxonomy Meets Plant Counting: A Fine-Grained, Taxonomic Dataset for Counting Hundreds of Plant Species — Xu et al. (2026) (arXiv:2603.21229, 2026)

What this evaluates

Evaluates class-agnostic counting (CAC) capabilities on fine-grained, taxonomically diverse plant species under few-shot exemplar guidance. It probes models' ability to handle dense occlusion, multi-scale observation, and cross-domain generalization from generic objects to biological imagery.

Datasets

• TPC–268 — total 10000; splits: train (-1), val (-1), test (-1); repo https://github.com/tiny-smart/TPC-268

Metrics

• MAE (primary) — range: other
  • Mean Absolute Error: the average of the absolute differences between predicted and ground-truth counts across all test images.
• RMSE — range: other
  • Root Mean Square Error: the square root of the average of squared differences between predicted and ground-truth counts.
• R^2 — range: other
  • Coefficient of Determination: measures the proportion of variance in the true counts explained by the model predictions.

Input / output format

Input: An image containing plants, accompanied by K (1 or 3) exemplar patches cropped from the same image. Optionally, text prompts specifying species name or full taxonomic hierarchy.

Output: Predicted count per image (via density map regression or bounding box detection).

Scoring recipe

def compute_metrics(pred_counts, true_counts):
    errors = pred_counts - true_counts
    mae = np.mean(np.abs(errors))
    rmse = np.sqrt(np.mean(errors**2))
    ss_res = np.sum(errors**2)
    ss_tot = np.sum((true_counts - np.mean(true_counts))**2)
    r2 = 1 - (ss_res / ss_tot) if ss_tot > 0 else 0.0
    return mae, rmse, r2

Common pitfalls

• Regression-based density estimation often outperforms detection-based bounding box approaches due to severe occlusion and structural entanglement in plant imagery.
• Global attention models (e.g., CACViT, TasselNetV4) tend to overfit validation splits during hyperparameter tuning, leading to significant test-set generalization gaps.
• Text-only or taxonomy-only prompts without visual exemplars yield inferior counting accuracy compared to visual-exemplar baselines.

Evidence (verbatim from paper)

A key observation is the substantial performance degradation across all models when trained on FSC–147 and tested on TPC–268, indicating that the model trained on generic objects struggle to generalize to plants. Conversely, when the training and testing datasets are exchanged, the MAE shows negligible increases or even reductions.

Citation

@misc{xu2026planttaxonomy,
  title={Plant Taxonomy Meets Plant Counting: A Fine-Grained, Taxonomic Dataset for Counting Hundreds of Plant Species},
  author={Xu et al. (2026)},
  year={2026},
  note={arXiv:2603.21229}
}

• arXiv: 2603.21229