Deep Learning for Computer Vision End Term: 13 April 2025, Set 1-4 (January 2025 term)

The CNN backbone used in DETR has a fixed receptive field that’s too large forsmall objects.

The global self-attention mechanism in DETR tends to dilute the signal fromsmall objects across all image locations.

DETR’s loss function actively discards all detections smaller than 50 × 50 pixels.

The object queries in DETR are programmed to ignore any object smaller than10% of the image size.

DETR’s encoder-decoder architecture was specifically designed to only detectobjects larger than a cat.

Linear contrast stretching is a point operation.

Moving average is an example of a local operation.

Convolution in the spatial domain can be obtained through addition in thefrequency domain.

All of these.

Image patching → Positional embedding → Linear projection of flattenedpatches → Transformer encoder→ Classification head

Image patching → Linear projection of flattened patches → Positionalembedding → Transformer encoder→ Classification head

Positional embedding → Image patching → Linear projection of flattenedpatches → Transformer encoder→ Classification head

Linear projection of Images → Image patching → Positional embedding →Transformer encoder → Classification head

Linear projection of Images → Image patching → Positional embedding →Transformer encoder → Transformer decoder→ Classification head