D-FINE: a better way to draw the box.

A bounding box often arrives as four numbers, which makes localization look like four ordinary regression tasks. But a partially occluded or blurry edge can be ambiguous. What if the model represented plausible positions as a distribution before producing a coordinate?

The idea behind D-FINE

D-FINE reframes box regression in DETR-based detection through fine-grained distribution refinement (FDR). Decoder layers iteratively refine localization distributions. Its second component, global optimal localization self-distillation (GO-LSD), transfers localization knowledge from stronger predictions within the model during training.

This note isolates the idea of decoding a distribution. It is not a reimplementation of the paper’s FDR parameterization, weighting function, or training losses.

FIG. 003 — An illustrative broad distribution becomes more concentrated after refinement. This is a conceptual drawing, not measured model output.

A coordinate as an expectation

For an educational example, choose candidate positions $b_0,\ldots,b_{B-1}$ and predict a logit for each. After softmax, the probabilities define a weighted mean:

$$ p_j = \frac{e^{z_j}}{\sum_k e^{z_k}}, \qquad \hat{b} = \sum_{j=0}^{B-1} p_j b_j. $$

READ THE EQUATION

Turn candidate-position logits into probabilities, then take their probability-weighted coordinate.

Term	What it is and does	What it controls
$z_j,z_k$	Raw scores for candidate positions; j selects one, k runs over all candidates.	Which candidate locations the model favors.
$e^{z_j}$, $\sum_k$, fraction bar	Exponentiate scores and divide by total evidence.	Produces nonnegative probabilities summing to one.
$p_j$	Probability of candidate bin j.	Its contribution to the predicted coordinate.
$b_j$, $B$, $j=0,\ldots,B-1$	Candidate coordinate, number of bins, and zero-based bin index.	The support and resolution available to the prediction.
$\hat b$, $\sum p_jb_j$	Predicted coordinate: multiply each candidate by its probability and add.	A continuous weighted expectation over the candidate coordinates.

Check: Candidates [0, 0.5, 1] with probabilities [0.2, 0.3, 0.5] give coordinate 0.65. Concentrating all mass on the middle bin gives 0.5. This is a teaching example, not a complete D-FINE implementation.

The coordinate remains differentiable with respect to the logits. Neighboring bins can share probability, allowing the expected coordinate to lie between them.

python
import torch

# A toy edge-position distribution, not D-FINE's full decoder.
bins = torch.linspace(0.0, 1.0, steps=5)
logits = torch.tensor([[-2.0, -0.5, 2.0, 1.0, -2.0]])
probabilities = logits.softmax(dim=-1)
coordinate = (probabilities * bins).sum(dim=-1)

assert coordinate.shape == (1,)
assert 0.0 <= coordinate.item() <= 1.0
print(round(coordinate.item(), 3))  # 0.548

Open complete file

Try increasing the fourth logit. The expected edge moves to the right. Try making all logits equal: the estimate moves to the center, despite the broad distribution.

An intuitive toy refinement can add a residual to the logits:

$$ z^{(\ell+1)} = z^{(\ell)} + \Delta z^{(\ell)}. $$

READ THE EQUATION

Form the next layer’s logits by adding a predicted correction to the current logits.

Term	What it is and does	What it controls
$z^{(\ell)}$, $z^{(\ell+1)}$	Current and next logit vectors. Superscripts in parentheses denote layer indices, not powers.	Which refinement stage is being represented.
$\ell$	Layer index.	The step through the refinement process.
$\Delta z^{(\ell)}$	A correction vector predicted at the current stage. Delta means change.	How each candidate’s relative score is revised.
$+$	Elementwise residual addition.	Preserves the current vector while allowing a correction of either sign.

Check: Starting logits [0, 0, 0] and correction [0, 0, 2] produce [0, 0, 2], favoring the last bin after softmax. A zero correction leaves the logits unchanged.

That update can strengthen some candidate positions and weaken others. A distribution contains more intermediate information than its mean alone; two distributions with the same mean can have very different shapes.

Reading the paper with a purpose

Follow one predicted edge through the model. Identify its reference geometry, the positions or offsets represented by the bins, the refinement update, and the final box decoding. Then look at which predictions act as teachers in self-distillation and where gradients are stopped.

This turns a crowded architecture diagram into a concrete data path you can reason about.

D-FINE: a better way to draw the box.

The idea behind D-FINE

A coordinate as an expectation

Refinement across layers

Reading the paper with a purpose

Further reading

Term	What it is and does	What it controls
\(z_j,z_k\)	Raw scores for candidate positions; j selects one, k runs over all candidates.	Which candidate locations the model favors.
\(e^{z_j}\), \(\sum_k\), fraction bar	Exponentiate scores and divide by total evidence.	Produces nonnegative probabilities summing to one.
\(p_j\)	Probability of candidate bin j.	Its contribution to the predicted coordinate.
\(b_j\), \(B\), \(j=0,\ldots,B-1\)	Candidate coordinate, number of bins, and zero-based bin index.	The support and resolution available to the prediction.
\(\hat b\), \(\sum p_jb_j\)	Predicted coordinate: multiply each candidate by its probability and add.	A continuous weighted expectation over the candidate coordinates.

Term	What it is and does	What it controls
\(z^{(\ell)}\), \(z^{(\ell+1)}\)	Current and next logit vectors. Superscripts in parentheses denote layer indices, not powers.	Which refinement stage is being represented.
\(\ell\)	Layer index.	The step through the refinement process.
\(\Delta z^{(\ell)}\)	A correction vector predicted at the current stage. Delta means change.	How each candidate’s relative score is revised.
\(+\)	Elementwise residual addition.	Preserves the current vector while allowing a correction of either sign.