forward method

Map<String, List<ValueVector>> forward(ValueVector backboneFeature)

Forward pass for the object detection head.

Takes a single ValueVector representing the aggregated image feature (e.g., the CLS token output from the ViT backbone).

Returns a Map containing:

• 'boxes': List of numQueries ValueVectors, each of 4 bounding box coordinates
• 'logits': List of numQueries ValueVectors, each of (numClasses + 1) class logits
• 'embeddings': List of numQueries ValueVectors, each of embeddingDim dimensions

Implementation

Map<String, List<ValueVector>> forward(ValueVector backboneFeature) {
  // Predict flattened bounding box coordinates for all queries
  final ValueVector rawBboxesFlat =
      bboxRegressionHead.forward(backboneFeature);

  // Predict flattened class logits for all queries
  final ValueVector classLogitsFlat =
      classPredictionHead.forward(backboneFeature);

  // NEW: Predict flattened face embeddings for all queries
  final ValueVector faceEmbeddingsFlat =
      faceEmbeddingHead.forward(backboneFeature);

  // Reshape flattened outputs into lists of ValueVectors for each query
  final List<ValueVector> predictedBoxes = [];
  for (int i = 0; i < numQueries; i++) {
    predictedBoxes.add(ValueVector(rawBboxesFlat.values
        .sublist(i * numBoxCoords, (i + 1) * numBoxCoords)));
  }

  final List<ValueVector> predictedLogits = [];
  for (int i = 0; i < numQueries; i++) {
    predictedLogits.add(ValueVector(classLogitsFlat.values
        .sublist(i * (numClasses + 1), (i + 1) * (numClasses + 1))));
  }

  // NEW: Reshape flattened embeddings
  final List<ValueVector> predictedEmbeddings = [];
  for (int i = 0; i < numQueries; i++) {
    predictedEmbeddings.add(ValueVector(faceEmbeddingsFlat.values
        .sublist(i * embeddingDim, (i + 1) * embeddingDim)));
  }

  return {
    'boxes': predictedBoxes,
    'logits': predictedLogits,
    'embeddings': predictedEmbeddings, // Add new output
  };
}