MegaDetector Pipeline Architecture

Overview

The MegaDetector pipeline processes large datasets of camera trap images, detects objects (animals, humans, vehicles), and stores results in designated Azure Blob Storage containers.

Key Components

  1. Datastores:
    • Deployed via Terraform/Terragrunt as part of the infrastructure-as-code setup.
    • Refer to the NIP-Lakehouse-Infra repo for the Terraform code used to deploy these datastores
    • landing_kutuma_hashed: Stores the raw camera trap images.
    • ml_public_models: Holds the MegaDetector model (md_v5b.0.0.pt).
    • bronze_camera_trap: Stores intermediate processed results.
    • bronze_megadetector: Stores the final detection results.
  2. Pipeline Phases:
    • Phase 1: Image ingestion and initial object detection.
      • The script cameratraps_pipeline.py is executed, and results are stored in bronze_camera_trap.
    • Phase 2 & 3: Post-processing and archiving of the images. The final results are saved in bronze_megadetector.
  3. Pipeline Details:
    • Pipeline Name: MegaDetector-NaturalState-RBP-Pipeline
    • Inference Script: cameratraps_pipeline.py
    • Model Used: md_v5b.0.0.pt
    • Execution Environment: megadetector-env (version 4)
    • Compute Clusters:
      • GPU Cluster: Used for Phase 1 (object detection).
      • CPU Cluster: Used for Phases 2 & 3 (post-processing and archiving).