Skip to content

Detection-Only Model Workflow

Use this guide when you already have a YOLO detection model and want to use IntegraPose for inference, ROI-aware analytics, and batch processing.

What this workflow is good for

Good fit Less ideal
Arena occupancy, entries/exits, dwell time, object-zone summaries Pose-driven latent modeling in Tab 7
Detection-based event timing Fine-grained posture analysis
Large multi-video inference and review workflows Training a detection model inside Tab 3

Important limitation up front

IntegraPose can run detect inference and analyze detection-only outputs in Bout Analytics, but the built-in Model Training tab is pose-oriented and Behavior Clustering (Tab 7) requires pose data.

If you only have detection outputs, your usual stopping point is Bout Analytics.

Existing detection model
  -> Tab 4 Inference (task = detect)
  -> Tab 6 Bout Analytics
  -> Optional Batch Processing Wizard / plugins

Step 1. Prepare videos

Use Data Preprocessing if you need to:

  • extract frames from raw videos
  • crop videos to a shared arena
  • clean up input folders before running inference

If your videos are already ready to process, you can skip directly to Inference.

Step 2. Run file inference in Tab 4

Use Inference with these settings:

Setting Recommendation
Model artifact Your existing YOLO detection checkpoint
Inference task detect or auto when the model name clearly indicates detection
Save Results (.txt) On
Use Tracker On for multi-animal work, optional for single-animal work
Project / Run Name Set these if you want predictable output folders

Detection workflow notes

  • Detection-only outputs still work in Bout Analytics.
  • Tracking is strongly recommended for multi-animal recordings.
  • Webcam Inference is more naturally aligned with pose workflows, so start with file inference unless you have a very specific live detection need.

Step 3. Analyze behavior in Tab 6

Open Bout Analytics and set:

Input What to provide
Source Video The video used for inference
YOLO Output Folder The folder containing the detection .txt outputs
ROI settings Your arena zones and thresholds

What works with detection-only labels

  • Bout timing
  • ROI entries and exits
  • Dwell time
  • Zone transitions
  • Reviewed bout exports
  • Batch analytics output manifests

What changes without pose keypoints

  • ROI logic falls back to box and center evidence rather than body-part entry logic.
  • Keypoint-dependent metrics are unavailable or less informative.
  • Tab 7 does not become a full downstream modeling path from detection-only data.

Step 4. Use batch processing when scale matters

If you have many recordings:

  1. Open File -> Batch Processing Wizard...
  2. Queue videos
  3. Reuse shared ROI settings when appropriate
  4. Run inference + analytics in one pass
  5. Review selected videos or exports after completion

See the Batch Processing Wizard guide for the detailed UI.

Step 5. Optional tools

You may also find these useful:

Tool When to use it
Zone Counter plugin Live zone counts during active experiments
EDA Tool plugin Explore output tables and quick plots
TandemYTC - Tandem YOLO + Temporal Classifier plugin Temporal classification workflows outside Tab 7

See the Plugin Catalog for the full list.

Detection-only checklist

  • Use a valid detection checkpoint
  • Save YOLO text results
  • Enable tracking for multi-animal recordings
  • Point Bout Analytics to the same video and output folder
  • Expect Tab 6 to be the main endpoint for analysis