# Agentic Ecology Camera Traps

> Provides capabilities to run SpeciesNet detector and classifier on camera trap images, extract crop-level feature embeddings, and populate a Hoplite vector database for downstream search and agile modeling. Use when ingesting camera trap images, classifying wildlife photos, or indexing camera trap data.

- Skill: `google-deepmind/agentic-ecology-camera-traps` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add google-deepmind/agentic-ecology-camera-traps`
- Raw SKILL.md: https://api.skillmd.com/api/skills/google-deepmind/agentic-ecology-camera-traps/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- License: Apache-2.0
- Author: google-deepmind (https://skillmd.com/u/google-deepmind)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/google-deepmind/agentic-ecology-camera-traps

---


# Camera Traps Skill

Use this skill when you need to process a collection of camera trap images, run
species classification, extract vector representation embeddings, and store them
inside a Hoplite vector database.

## Workflow Overview

Follow these sequential steps:

1. **Identify Dataset and Limits**:
   - Locate the target camera trap images directory.
   - Assess if a GPU is available on the system. If running on CPU-only,
     discuss with the user or apply a processing limit (e.g., first 1000
     images) to prevent the ingestion pipeline from running excessively long.
1. **Initialize Hoplite Database**:
   - Create a Hoplite database (`SQLiteUSearchDB`) at the destination folder.
   - Configure it with an embedding dimension of `1280` (EfficientNet-V2 M
     feature size), the metric set to `Cos`, and the data type set to
     `float16`.
1. **Run Ingestion Pipeline**:
   - Instantiate the `SpeciesNetDetector` and `SpeciesNetClassifier` models
     (if running in an environment with pre-mounted read-only models like
     `/kaggle/input/` on Colab, copy the model directory to a local writable
     path first; see the [Technical Reference](references/REFERENCE.md)).
   - Register a PyTorch forward hook on the classifier's average pooling
     layer (`SpeciesNet/efficientnetv2-m/avg_pool/Mean_Squeeze__3825`) to
     intercept raw embeddings.
   - For each image:
     - Insert it into the database as a recording.
     - Run the detector model to get bounding box coords for animal
       detections.
     - Crop the PIL image to the bounding box, preprocess it, and run the
       classifier to extract the 1280-dim embedding vector.
     - Cast the vector to `float16` and insert it into the database as a
       window.
   - Commit database changes periodically and at the end of ingestion via
     `db.commit()` to persist SQLite rows and flush the USearch vector index to disk.
1. **Agile Modeling and Search**:
   - Once populated, use the Hoplite database to perform vector searches
     (ranking by similarity) or train active learning classifiers on top of
     the embeddings.
1. **Camera Trap Visualization Guidelines (M3 UI)**:
   - **Context Preservation**: Avoid displaying raw cropped images in the
     result cards. Instead, display the original (uncropped) image inside the
     card container and draw the animal detection as a red border box overlay
     dynamically using CSS absolute positioning and percentages (e.g., `left: xmin * 100%`, `top: ymin * 100%`, etc.).
   - **Full-Resolution Modal Preview**: Implement a click handler on the card
     media that triggers a floating fullscreen modal containing the uncropped
     image and the aligned bounding box overlay to allow the user to verify
     low-confidence detections.
   - **Custom Query Search Support**: The backend server supporting the Web
     UI must implement on-the-fly embedding extraction for custom HTTP/S
     query URIs by downloading the image, running the detector to identify
     target bounding boxes, preprocessing the crop, and capturing the
     embedding vector using the PyTorch forward hook on the classifier.

## Technical Reference

For detailed model loading code, Kaggle/Colab read-only filesystem workarounds,
PyTorch hook embedding extraction, and bounding box calculations, see:

- [Camera Traps Technical Reference](references/REFERENCE.md)