Ultralytics YOLOv8 - Open Images V7

Mix.install(
  [
    {:yolo, ">= 0.0.0"},
    {:nx, "~> 0.9"},
    {:exla, "~> 0.9"},
    {:image, "~> 0.54"},
    {:evision, "~> 0.2"},
    {:kino, "~> 0.16"},
    {:kino_yolo, github: "poeticoding/kino_yolo", branch: "main"},
    {:pythonx, "~> 0.4.2"},
    {:kino_pythonx, "~> 0.1.0"}
  ],
  config: [
    nx: [default_backend: EXLA.Backend]
  ]
)

Application.put_env(:ortex, Ortex.Native, features: [:cpu])

[project]
name = "ultralytics_to_onnx"
version = "0.0.0"
requires-python = "==3.13.*"
dependencies = [
  "ultralytics==8.3.155",
  "onnx==1.18.0",
  "onnxruntime==1.22.0",
  "onnx_pytorch==0.1.5",
  "requests"
]

Ultralytics yolov8l-oiv7 to ONNX

# https://docs.ultralytics.com/datasets/detect/open-images-v7/
model_name = "yolov8x-oiv7"

image_path = "benchmarks/images/table.jpg"
model_path = "#{model_name}.onnx"
classes_path = "#{model_name}_classes.json"
:ok

:ok

Pythonx.eval(
  """
  from ultralytics import YOLO
  import json
  
  IMAGE_SIZE = 640
  model_name = model_name.decode("utf-8")
  
  model = YOLO(model_name)
  model.export(format='onnx', imgsz=IMAGE_SIZE, opset=12)

  # Export the categories
  with open(classes_path, "w") as f:
    data = [model.names[idx] for idx in model.names]
    json.dump(data, f)
  """,
  %{"model_name" => model_name, "classes_path" => classes_path}
)
:ok

nil

Load and Run

mat = Evision.imread(image_path)

%Evision.Mat{
  channels: 3,
  dims: 2,
  type: {:u, 8},
  raw_type: 16,
  shape: {1138, 640, 3},
  ref: #Reference<0.2823733541.3473539090.13647>
}

model = YOLO.load(model_path: model_path, classes_path: classes_path, eps: [:cpu])

detected_objects = 
  model
  |> YOLO.detect(mat, iou_threshold: 0.45, prob_threshold: 0.25)
  |> YOLO.to_detected_objects(model.classes)

18:27:35.461 [info] Loaded model yolov8x-oiv7.onnx with [:cpu] execution providers

18:27:35.461 [info] Initialized model yolov8x-oiv7.onnx

[
  %{
    class: "Food",
    prob: 0.2534943222999573,
    bbox: %{h: 38, w: 73, cx: 257, cy: 468},
    class_idx: 199
  },
  %{
    class: "Food",
    prob: 0.3215067982673645,
    bbox: %{h: 72, w: 134, cx: 512, cy: 515},
    class_idx: 199
  },
  %{
    class: "Food",
    prob: 0.35645782947540283,
    bbox: %{h: 35, w: 99, cx: 351, cy: 385},
    class_idx: 199
  },
  %{
    class: "Food",
    prob: 0.5081610679626465,
    bbox: %{h: 70, w: 117, cx: 389, cy: 484},
    class_idx: 199
  },
  %{
    class: "Kitchen appliance",
    prob: 0.2527419328689575,
    bbox: %{h: 45, w: 64, cx: 38, cy: 33},
    class_idx: 291
  },
  %{
    class: "Table",
    prob: 0.7417454123497009,
    bbox: %{h: 896, w: 640, cx: 320, cy: 673},
    class_idx: 514
  }
]

{:ok, image} = Image.from_evision(mat)
KinoYOLO.Draw.draw_detected_objects(image, detected_objects)