DN-DETR Learns by Undoing Noise

DN-DETR feeds noised truth boxes with labels to the decoder, so queries learn to denoise and matching stops flipping.

Five denoising groups with 0.04 noise steady matching in 12 epochs

Why Does This Exist?

Bipartite matching has no anchor prior, so early epochs assign one object to many queries in turn. Gradients fight and loss plateaus.

Denoising gives a stable bypass. It builds on detr, precedes dino contrast, and aids object detection training.

Think of It Like This

Copying with smudged answer sheets

You get smudged copies of true answers plus blank sheets. Rebuilding smudged copies teaches spelling while blanks learn layout.

Noised queries are smudged copies, learnable queries are blanks. Where it stops: smudges too faint teach copying, not finding, just as tiny noise adds no signal.

How It Actually Works

For each truth $(c,b)$ , make $G = 5$ noised copies with box jitter $\lambda = 0.4$ scaled to $0.04$ absolute plus $20$ percent label flips. Mask attention so denoising and matching queries never see each other. Loss sums reconstruction plus standard set loss.

Worked example

Truth $w = 0.20$ . Noised $0.24$ has error $0.04$ . Model rebuilds $0.21$ , error $0.01$ , a four times cut. With $2$ truths and $5$ groups, $10$ denoising queries join $100$ matching queries. Matching instability (flips per epoch) drops from $9$ to $2$ by epoch $12$ in paper runs.

Watch Out For

Leaking denoising queries into matching attention

Unmasked denoising lets matching queries copy truths. Symptom: train loss dives but val AP stalls. Fix: keep the attention mask that blocks denoising to matching visibility.

The Quick Version

DN-DETR adds noised truth queries beside learnable matching queries.
Box noise 0.04 plus label flips teach rebuild plus classify.
Fixture error 0.04 to 0.01 shows four times denoising gain.
Five groups add ten queries to one hundred matching ones.
Mask groups apart or matching leaks and val stalls.