DETRConfig

Configuration for DETR (DEtection TRansformer).

DETR (Carion et al., ECCV 2020) reformulates object detection as a direct set prediction problem. A CNN backbone extracts image features, a Transformer encoder-decoder processes them with N learned object queries, and FFN heads predict class labels + normalised (cx, cy, w, h) boxes. Training uses the Hungarian algorithm to match predictions to ground-truth objects (no hand-crafted anchors, no NMS at inference).

Architecture overview: Image → ResNet-50 (C5 feature map) → 1×1 projection (d_model) → flatten + 2-D sinusoidal positional encoding → Transformer encoder (N_enc layers, d_model, n_head) → Transformer decoder (N_dec layers, N_queries object queries) → FFN: cls head (num_classes + 1) + box head (4, sigmoid)

Args: num_classes: Number of foreground classes (background = class 0). in_channels: Input image channels.

-- Backbone -- backbone_layers: ResNet-50 layer counts (default 3,4,6,3).

-- Transformer -- d_model: Transformer embedding dimension. n_head: Number of self-/cross-attention heads. num_encoder_layers: Encoder depth. aux_loss: Apply the set loss after every decoder layer, not only the last (paper: "auxiliary decoding losses"). num_decoder_layers: Decoder depth. dim_feedforward: FFN inner dimension in each transformer layer. dropout: Dropout in transformer. num_queries: Number of object queries (N).

-- FFN heads -- num_bbox_layers: MLP depth for bounding-box head. bbox_hidden_dim: Hidden width inside the bbox MLP.

-- Inference -- score_thresh: Minimum predicted class probability (before NMS).