Skip to content
finwave Docs

Pre-Training

What you’ll learn

Section titled “What you’ll learn”
  • What pre-training is and when to use it
  • How the desktop client extracts training crops from your images
  • How the three-tier detection system classifies training data quality
  • How to review the training data quality report
  • How to upload training data to start model training

What is pre-training?

Section titled “What is pre-training?”

Pre-training is an optional step you can run before onboarding. If you have a collection of labeled historical images — photos where you already know which individual animal is in each image — you can use pre-training to build an identification model for your population before uploading your full dataset.

This means that when your encounters are onboarded, the ML pipeline already has a trained ID model ready to suggest identifications, rather than starting from scratch.

How training data extraction works

Section titled “How training data extraction works”

The desktop client uses a local YOLOv8 object detection model to find features of interest (dorsal fins, eye patches, or other markers) in your labeled images. This detection runs entirely on your machine using ONNX Runtime on CPU — no GPU is required, and no images are sent to the server during this step.

For each image, the detector produces zero or more bounding box detections. The client then crops each detection from the original image and pairs it with the individual ID label you provided. These cropped images become the training data that will later be uploaded to finwave’s training pipeline.

Crops are stored locally at ~/.finwave/populations/{population-id}/training-data/.

The tier system

Section titled “The tier system”

Not every image produces a clean, unambiguous training sample. The desktop client classifies each image into one of three tiers based on how many detections were found and how confident the detector is:

Tier 1 — Clean pairs

Section titled “Tier 1 — Clean pairs”

A single detection is found in the image. Since only one feature of interest is present, it maps directly to the labeled individual ID. These are high-confidence training samples.

Tier 2 — Prominent detection

Section titled “Tier 2 — Prominent detection”

Multiple detections are found, but one is clearly dominant. The client scores each detection using a combination of bounding box area, proximity to image center, and model confidence. If the top-scoring detection exceeds the prominence threshold (default: 0.7), it is selected as the training sample. The remaining detections are discarded.

Tier 3 — Ambiguous

Section titled “Tier 3 — Ambiguous”

Multiple detections are found with no clearly dominant one, or no detection meets the prominence threshold. These images are skipped entirely and do not contribute to training data.

The training data quality report

Section titled “The training data quality report”

After extraction completes, the desktop client shows a quality report summarizing your training data:

  • Tier breakdown — How many images fell into each tier, shown as counts and percentages
  • Unique individuals — The number of distinct individual IDs in your training set
  • Samples per individual — A distribution showing how many training crops each individual has, highlighting individuals with very few samples

Review this report before uploading. A healthy training set has most images in Tier 1 or Tier 2, covers many unique individuals, and has at least several samples per individual.

Uploading training data

Section titled “Uploading training data”

Once you are satisfied with the quality report, you upload the training data to finwave. The server queues a model training job for your population. Training runs on finwave’s infrastructure, and you will receive a notification when the model is deployed and ready.

After the model is deployed, any encounters you onboard will benefit from ID predictions immediately.

Section titled “Related”