projects/drones.yaml

# =====================================================================
# drones.yaml — example project config for the data labeling factory
# =====================================================================
#
# This is the canonical example. It captures EXACTLY what we built tonight
# for the fiber-optic drone detector. To make a new project (e.g. stop signs,
# fire hydrants, manufacturing defects), copy this file, change `target_object`,
# adjust the queries, and run:
#
#     data_label_factory pipeline --project projects/drones.yaml
#
# Generic shape:
#   project_name      → human-readable identifier (used in experiment dir names)
#   target_object     → the thing you're trying to detect (templated into prompts)
#   data_root         → where local images go
#   r2:               → cloud storage config (bucket, prefix per stage)
#   buckets:          → gather plan (5 buckets is conventional but any structure works)
#   falcon_queries:   → list of queries to run Falcon Perception with
#   prompts:          → optional overrides for templated prompts
#   backends:         → which model backend to use per stage
# =====================================================================

project_name: drones
target_object: "fiber optic drone"
description: |
  Auto-labeling pipeline for fiber-optic drone detection. Falcon Perception
  grounds bboxes for any drone, spool, or cable; Qwen2.5-VL verifies each.

# Where local images live (gitignored)
data_root: ~/drone-falcon-data/v2

# Cloudflare R2 storage
r2:
  bucket: drone-falcon
  raw_prefix: raw_v2/                    # gathered images
  labels_prefix: labels/                 # COCO + verified JSONs
  reviews_prefix: labels/reviews.json    # human verdicts saved by web UI

# What to gather, organized by bucket. Each bucket is a folder under data_root
# and a corresponding R2 prefix. Multiple queries are OR'd via DDG/Wikimedia.
buckets:
  positive/fiber_spool_drone:
    queries:
      - "fiber optic FPV drone"
      - "tethered fiber optic drone"
      - "Ukraine fiber optic drone war"
      - "fiber optic kamikaze drone"
      - "fiber optic drone payload"
      - "wired FPV drone Ukraine"
      - "fiber optic drone with spool"
      - "Russian fiber optic drone"
      - "fiber optic dispenser drone"
      - "fiber optic combat drone"

  positive/spool_only:
    queries:
      - "fiber optic cable spool"
      - "optical fiber reel"
      - "fiber optic winding machine"
      - "spooled optical fiber cable"
      - "fiber optic cable on reel"
      - "optical fiber cable drum"

  negative/drones_no_spool:
    queries:
      - "DJI Mavic 3 Pro photo"
      - "FPV racing drone closeup"
      - "consumer quadcopter flying"
      - "agricultural spraying drone"
      - "DJI Mini 4 Pro photo"
      - "Autel Evo drone"
      - "Skydio 2 drone"
      - "racing drone build"

  distractor/round_things:
    queries:
      - "garden hose reel"
      - "cable drum reel industrial"
      - "duct tape roll"
      - "fire hose reel"
      - "rope coil pile"
      - "extension cord reel"
      - "thread spool sewing"

  background/empty:
    queries:
      - "blue sky clouds"
      - "open field landscape"
      - "industrial workshop interior"
      - "outdoor military training"

# What Falcon Perception should look for in each image (the bbox grounding queries).
# These are the specific objects we want bounding boxes on.
falcon_queries:
  - "fiber optic spool"
  - "cable spool"
  - "drone"
  - "quadcopter"
  - "fiber optic drone"

# Optional: prompt overrides. Default templates use {target_object} substitution.
# Leave commented to use the defaults from lib/project.py.
prompts:
  filter: |
    Look at this image. Does it show a {target_object}, a related component
    (cable spool, fiber reel, wound cable), or any other relevant object?
    Answer with exactly one word: YES or NO.
    YES if you see ANY of: a {target_object}, a quadcopter, a cable reel, a fiber spool.
    NO if the main subject is something else.

  verify: |
    Look carefully at this image crop.
    Question: Is the main object in this crop actually a {query}?
    Answer first with one word: YES, NO, or UNSURE.
    Then briefly say what the object actually is in 5-10 words.

# Which model backend to use per stage.
# qwen   = Qwen2.5-VL-3B via mlx-vlm server (M4 :8291)  — fast, free
# gemma  = Gemma 4 26B via mac_tensor (M4 :8500)        — slow chained agent
# falcon = Falcon Perception via mac_tensor /api/falcon — bbox grounding only
# pod    = remote RunPod GPU pod                        — fast, ~$0.15-1.65/run
backends:
  filter: qwen
  label:  pod          # Falcon on RunPod for production scale
  verify: pod          # Qwen on the same pod

# Pod settings (only used when a stage backend = pod)
pod:
  gpu_types: ["NVIDIA L40S"]
  data_centers: ["EU-RO-1", "EU-CZ-1", "EU-NL-1", "US-CA-2"]
  image: "runpod/pytorch:1.0.3-cu1290-torch291-ubuntu2204"
  container_disk_gb: 30
  volume_gb: 30