Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Hierarchical Re-Classification: Combining Animal Classification Models with Vision Transformers

Hugo Markoff, Jevgenijs Galaktionovs

TL;DR

The paper tackles the challenge of conservative, high-level taxonomic predictions in wildlife classification by introducing a hierarchical re-classification system that fuses SpeciesNet outputs, CLIP embeddings, and metric learning within a five-stage pipeline. By anchoring on high-confidence predictions and refining through centroid clustering, triplet-loss embeddings, and adaptive scoring, the approach achieves 96.5% accuracy on re-classified detections and 64.9% species-level upgrades on 4,018 images from the Desert Lion dataset. This yields substantial reductions in manual annotation burden and demonstrates production-ready integration within the Animal Detect platform, with strong potential for generalization to hierarchical taxonomies in related domains. The work also outlines concrete avenues for improvement, including more principled scoring, hierarchical processing by taxonomic level, and broader benchmarking to support open, reproducible comparisons.

Abstract

State-of-the-art animal classification models like SpeciesNet provide predictions across thousands of species but use conservative rollup strategies, resulting in many animals labeled at high taxonomic levels rather than species. We present a hierarchical re-classification system for the Animal Detect platform that combines SpeciesNet EfficientNetV2-M predictions with CLIP embeddings and metric learning to refine high-level taxonomic labels toward species-level identification. Our five-stage pipeline (high-confidence acceptance, bird override, centroid building, triplet-loss metric learning, and adaptive cosine-distance scoring) is evaluated on a segment of the LILA BC Desert Lion Conservation dataset (4,018 images, 15,031 detections). After recovering 761 bird detections from "blank" and "animal" labels, we re-classify 456 detections labeled animal, mammal, or blank with 96.5% accuracy, achieving species-level identification for 64.9 percent

Hierarchical Re-Classification: Combining Animal Classification Models with Vision Transformers

TL;DR

The paper tackles the challenge of conservative, high-level taxonomic predictions in wildlife classification by introducing a hierarchical re-classification system that fuses SpeciesNet outputs, CLIP embeddings, and metric learning within a five-stage pipeline. By anchoring on high-confidence predictions and refining through centroid clustering, triplet-loss embeddings, and adaptive scoring, the approach achieves 96.5% accuracy on re-classified detections and 64.9% species-level upgrades on 4,018 images from the Desert Lion dataset. This yields substantial reductions in manual annotation burden and demonstrates production-ready integration within the Animal Detect platform, with strong potential for generalization to hierarchical taxonomies in related domains. The work also outlines concrete avenues for improvement, including more principled scoring, hierarchical processing by taxonomic level, and broader benchmarking to support open, reproducible comparisons.

Abstract

State-of-the-art animal classification models like SpeciesNet provide predictions across thousands of species but use conservative rollup strategies, resulting in many animals labeled at high taxonomic levels rather than species. We present a hierarchical re-classification system for the Animal Detect platform that combines SpeciesNet EfficientNetV2-M predictions with CLIP embeddings and metric learning to refine high-level taxonomic labels toward species-level identification. Our five-stage pipeline (high-confidence acceptance, bird override, centroid building, triplet-loss metric learning, and adaptive cosine-distance scoring) is evaluated on a segment of the LILA BC Desert Lion Conservation dataset (4,018 images, 15,031 detections). After recovering 761 bird detections from "blank" and "animal" labels, we re-classify 456 detections labeled animal, mammal, or blank with 96.5% accuracy, achieving species-level identification for 64.9 percent
Paper Structure (14 sections, 2 equations, 1 figure, 1 table)

This paper contains 14 sections, 2 equations, 1 figure, 1 table.

Table of Contents

  1. Introduction
  2. Background
  3. Method
  4. Pipeline Architecture
  5. SpeciesNet Ensemble Strategy
  6. Five-Stage Re-Classification
  7. Experiments
  8. Results
  9. Overall Performance
  10. Conservative Rollup Strategy
  11. User Interface Integration
  12. Discussion
  13. Planned Work
  14. Conclusion

Figures (1)

  • Figure 1: Pipeline: Detection → HITL → dual-path (SpeciesNet + CLIP) → 5-stage re-classification