Make it fast with YoloFastNMS

Mix.install([
  {:yolo, ">= 0.0.0"},
  {:yolo_fast_nms, ">= 0.0.0"},
  {:nx, "~> 0.9.1"},
  {:exla, "~> 0.9.0"},
  {:kino, "~> 0.14.0"},
  {:evision, "~> 0.2"}
],[
  config: [
    nx: [default_backend: EXLA.Backend]
  ]
])

Load image and model

mat = Evision.imread(Kino.FS.file_path("traffic.jpg"))

%Evision.Mat{
  channels: 3,
  dims: 2,
  type: {:u, 8},
  raw_type: 16,
  shape: {1080, 1920, 3},
  ref: #Reference<0.1364157824.1534197778.893>
}

model =
  YOLO.load(
    model_path: Kino.FS.file_path("yolov8n.onnx"),
    classes_path: Kino.FS.file_path("yolov8n_classes.json")
  )

%YOLO.Model{
  ref: #Ortex.Model<
    inputs: [
      {"images",
       "Tensor {\n    ty: Float32,\n    dimensions: [\n        1,\n        3,\n        640,\n        640,\n    ],\n}",
       [1, 3, 640, 640]}
    ]
    outputs: [
      {"output0",
       "Tensor {\n    ty: Float32,\n    dimensions: [\n        1,\n        84,\n        8400,\n    ],\n}",
       [1, 84, 8400]}
    ]>,
  classes: %{
    39 => "bottle",
    74 => "clock",
    59 => "bed",
    69 => "oven",
    67 => "cell phone",
    45 => "bowl",
    50 => "broccoli",
    22 => "zebra",
    51 => "carrot",
    26 => "handbag",
    63 => "laptop",
    47 => "apple",
    27 => "tie",
    77 => "teddy bear",
    0 => "person",
    5 => "bus",
    21 => "bear",
    62 => "tv",
    30 => "skis",
    16 => "dog",
    3 => "motorcycle",
    53 => "pizza",
    33 => "kite",
    14 => "bird",
    40 => "wine glass",
    37 => "surfboard",
    24 => "backpack",
    17 => "horse",
    48 => "sandwich",
    73 => "book",
    11 => "stop sign",
    57 => "couch",
    43 => "knife",
    6 => "train",
    20 => "elephant",
    60 => "dining table",
    28 => "suitcase",
    25 => "umbrella",
    1 => "bicycle",
    58 => "potted plant",
    32 => "sports ball",
    76 => "scissors",
    36 => "skateboard",
    35 => "baseball glove",
    15 => "cat",
    78 => "hair drier",
    64 => "mouse",
    75 => "vase",
    ...
  },
  model_impl: YOLO.Models.YoloV8,
  shapes: %{input: {1, 3, 640, 640}, output: {1, 84, 8400}}
}

YOLO.NMS vs YoloFastNMS

# YOLO.NMS
{nx_nms_time, _result} = :timer.tc fn ->
  YOLO.detect(model, mat)
end

# YOLO.NMS
{fast_nms_time, _result} = :timer.tc fn ->
  YOLO.detect(model, mat, nms_fun: &YoloFastNMS.run/3)
end

IO.puts("YOLO.NMS: #{trunc(nx_nms_time/1_000)}ms")
IO.puts("YoloFastNMS: #{trunc(fast_nms_time/1_000)}ms")

YOLO.NMS: 550ms
YoloFastNMS: 37ms

:ok