NudeNet

The detection process returns a result object containing raw parts and high-level composite classifications.

Result Structure

• input: Object containing { width, height } of the processed input.
• parts: An array of detected objects. Each object contains:
  • score: Confidence score (0-1).
  • id: The integer class ID.
  • class: The human-readable label (e.g., 'breast', 'vagina').
  • box: The bounding box in [x, y, width, height] format.
• person: Boolean, true if any part belonging to the person composite is detected.
• sexy: Boolean, true if any part belonging to the sexy composite is detected.
• nude: Boolean, true if any part belonging to the nude composite is detected.

Composite Definitions

NudeNet categorizes specific class IDs into three main groups:

• person: Includes IDs like female face and male face.
• sexy: Includes IDs like exposed armpits, belly, and feet.
• nude: Includes IDs like exposed anus, buttocks, and vagina.

To run NudeNet on a video stream in the browser, follow these steps:

1. Initialize TensorFlow.js: Set the backend to webgpu (if available) or webgl. Enable WEBGL_USE_SHAPES_UNIFORMS for performance.
2. Load the Model: Use tf.loadGraphModel(options.modelPath) to load the NudeNet graph model.
3. Process Frames:
   • Convert the video frame to a tensor using tf.browser.fromPixelsAsync(video).
   • Resize the tensor to the desired options.resolution.
   • Execute the model using model.executeAsync(t.batch, options.outputNodes).
   • Process the resulting tensors (boxes, scores, classes) through Non-Max Suppression (NMS).
4. Cleanup: Always use tf.dispose() on tensors to prevent memory leaks.

// Simplified logic flow for a detection loop
[t.boxes, t.scores, t.classes] = await model.executeAsync(t.batch, options.outputNodes);
const res = await processPrediction(t.boxes, t.scores, t.classes, t.cast);

// Cleanup tensors
Object.keys(t).forEach((tensor) => tf.dispose(t[tensor]));

When using NudeNet in a browser environment, you can control the detection behavior using an options object. Key configuration parameters include:

• modelPath: Path to the model.json file.
• minScore: Minimum confidence threshold for a detection (e.g., 0.38).
• maxResults: Maximum number of detections to return.
• iouThreshold: Intersection over Union threshold for Non-Max Suppression (NMS).
• outputNodes: Array of strings representing the model's output node names (e.g., ['output1', 'output2', 'output3']).
• resolution: A [number, number] tuple defining the input resolution for the model (e.g., [1280, 720]).
• blurRadius: Controls the intensity of the blur effect when applying redaction.

const options = {
  modelPath: '../models/default-f16/model.json',
  videoPath: '../samples/video.webm',
  minScore: 0.38,
  maxResults: 50,
  iouThreshold: 0.5,
  outputNodes: ['output1', 'output2', 'output3'],
  blurRadius: 25,
  resolution: [1280, 720] as [number, number],
};

NudeNet returns a structured object containing metadata about the input image and detection results. The results include boolean flags for high-level classifications (person, sexy, nude) and a detailed parts array for specific body part detections.

Output Object Schema:

• input: Object containing file (String), width (Number), and height (Number).
• person: Boolean indicating if a person was detected.
• sexy: Boolean indicating if the person is considered sexy.
• nude: Boolean indicating if the person is considered nude.
• parts: Array of detected body parts, where each part contains:
  • score: Confidence score (Number).
  • id: Detection ID (Number).
  • class: Label for the body part (String).
  • box: Bounding box coordinates as [x, y, width, height] (Number[]).

{
  input: {
    file: String,
    width: Number,
    height: Number,
  },
  person: Boolean,
  sexy: Boolean,
  nude: Boolean,
  parts: Array<{ 
    score: Number,
    id: Number,
    class: String,
    box: Number[],
  }],
}