Concept · 1 paper · updated 2026-11-15
Spatial propagation network
Refines a dense map by repeated weighted copying from neighbours. Fixed affinity weights (CSPN) are replaced by weights recomputed at every step (DySPN), with a diffusion-suppression term against over-smoothing. Author-reported on KITTI validation, DySPN at six steps beats CSPN at twelve. Only one card so far: DySPN, for depth completion.
computer-visiondepthaffinityrefinementattention
Markdown5 min read
AI-compiled overview, updated 2026-11-15, not yet reviewed by a human. Every claim cites its source; check it when a detail matters.
Overview
A spatial propagation network (SPN) refines a map by repeating one update many times. Each pixel takes a weighted combination of its neighbours, and the weights are called affinities. The repeated update spreads known values into unknown regions, guided by the image 2202.09769 §2. The property shared by the variants is that each step multiplies the map by a matrix whose rows total one 2202.09769 §3.1, eq. 1.
The design choice that separates the variants is when the affinities are computed. A fixed-affinity model computes them once and reuses them at every step. An adaptive model recomputes them at each step, so the weights can depend on how far the refinement has progressed 2202.09769 §1, §3.1. The paper also adds a stopping term, diffusion suppression, so that the update can converge early and stop blurring 2202.09769 §3.2.
Variants
Fixed-affinity linear propagation (CSPN)
The baseline that the DySPN paper compares against. Its affinity matrix stays fixed through all steps, and the paper argues that a 7×7 kernel places most of the weight in the outer ring, so more steps are needed to reach the same result 2202.09769 §1, §2. No card exists for this method, so the details here are only those the DySPN paper gives. Its reported 12-step error on KITTI validation is 789.6 mm RMSE 2202.09769 tab. 3.
Adaptive-affinity propagation, 7×7 window (DySPN)
Each of the six steps has its own attention maps, and these produce the affinity weights for that step 2202.09769 §3.3. The neighbourhood is the full 7×7 window, split into distance rings with one attention map per ring 2202.09769 eq. 9. This is the variant the paper uses for its ablation 2202.09769 tab. 3. Card: Dynamic Spatial Propagation Network for Depth Completion.
Dilated neighbourhood (DySPN)
Two 3×3 blocks, one with spacing 1 and one with spacing 3, replace the full window 2202.09769 eq. 10 2202.09769 §3.3. It uses fewer neighbours than the 7×7 window, and the paper reports the lowest per-module time among the three DySPN variants 2202.09769 tab. 4. Same card as above.
Deformable neighbourhood (DySPN)
A fixed 3×3 ring for the nearest neighbours, plus learned real-valued offsets for the outer distance rings 2202.09769 eq. 11 2202.09769 §3.3. This is the variant submitted to the KITTI depth completion benchmark, and it has the lowest validation error of the three 2202.09769 tab. 2, tab. 4. Same card as above.
Diffusion suppression (a component, not a variant)
A diagonal term that holds, for each pixel, the attention between the current and next depth. When the depth has converged, this term makes the step leave the map unchanged 2202.09769 §3.2, eq. 5, eq. 6. Removing it raises the 6-step 7×7 error on KITTI validation from 769.0 mm to 763.9 mm RMSE in the author's ablation 2202.09769 tab. 3. The gain is about 5 mm.
When to use which
The paper gives no cost model for other hardware. Compare the timings only within one table and one machine, since tab. 4 does not state either.
| Situation | Use | Why |
|---|---|---|
| Latency matters most and learned offsets are unwanted | Dilated neighbourhood | Lowest module time, 6.9 ms, at 748.5 mm validation RMSE 2202.09769 tab. 4 |
| Lowest error on KITTI and offsets are acceptable | Deformable neighbourhood | 739.4 mm validation RMSE, 10.3 ms 2202.09769 tab. 4; the benchmark submission at 709.12 mm RMSE 2202.09769 tab. 2 |
| A fixed, dense window is preferred and cost is less critical | 7×7 window | No learned offsets 2202.09769 eq. 9; 745.8 mm validation RMSE and 38 ms 2202.09769 tab. 4 |
| Few refinement steps must reach a given accuracy | Adaptive affinity (any DySPN variant) over fixed affinity | At six steps DySPN gives 769.0 mm against CSPN's 811.0 mm; CSPN at twelve steps gives 789.6 mm 2202.09769 tab. 3 |
| Indoor depth (NYU Depth v2) and the choice rests on cost | Any DySPN variant | The three variants fall within 0.090–0.091 m RMSE 2202.09769 tab. 1 |
| No training data or GPU is available | Not this family | All DySPN results need training on labelled depth; see depth-completion for a training-free option |
Our arithmetic from tab. 4: the three DySPN variants differ by about 9 mm in validation RMSE (739.4 to 748.5 mm), while their module times range from 6.9 to 38 ms. On KITTI validation the choice is therefore mostly a latency decision 2202.09769 tab. 4.
Implementation sketch
The shared core is one refinement loop. Variants differ only in which
neighbours k are read and where their attention maps come from.
# One refinement with N steps, adaptive affinity [arxiv:2202.09769 eq. 7]
# Inputs: depth0 (H, W) = initial depth h^0
# weight[k] (H, W): affinity for neighbour part k
# pi (N, 1 + K, H, W): attention per step; pi[t, 0] is the DS term
# Output: depth (H, W) after N steps
def refine(depth0, weight, pi, steps):
# Start from the initial depth [arxiv:2202.09769 §3.1]
depth = depth0
for t in range(steps):
# Signed and absolute normalisers [arxiv:2202.09769 eq. 8]
s = pi[t, 0] + sum(pi[t, k] * weight[k] for k in parts)
s_abs = pi[t, 0] + sum(pi[t, k] * abs(weight[k]) for k in parts)
# Weighted sum over neighbours, plus the centre term [eq. 7]
total = pi[t, 0] * depth + sum(
pi[t, k] * weight[k] * shift(depth, k) for k in parts
)
# Normalise, then pull back towards the initial depth [eq. 7]
depth = total / s_abs + (1 - s / s_abs) * depth0
return depthPoints the paper leaves open. Implementers must decide these themselves.
- Where the attention maps come from. The head outputs them with a sigmoid, but the channel layout is not given 2202.09769 §3.4.
- Zero normaliser. If every attention and weight in a neighbourhood is
zero,
s_absis zero. The paper has no guard 2202.09769 eq. 8. - Centre term without diffusion suppression. The paper does not define
pi[t, 0]when the term is switched off 2202.09769 §4.6. - Deformable sampling. Sampling at fractional offsets, such as bilinear interpolation, is not specified 2202.09769 eq. 11.
- Confidence map. The network predicts one, but no equation uses it 2202.09769 §3.4.
- Code. No official code is known. The DySPN card records this, and any reimplementation is unchecked against the authors' code.
Contradictions
No other card lists this concept, so no cross-paper disagreement is known. Within the DySPN paper, two points need care:
- Table 4 against Table 3. The 7×7 DySPN row in tab. 4 reports 745.8 mm validation RMSE at six steps. The matching row in tab. 3 with diffusion suppression reports 763.9 mm, and the sigmoid-only row reports 769.0 mm. The 745.8 mm value equals the stochastic-depth-only row of tab. 3, so tab. 4 probably uses that configuration, but the paper does not say so 2202.09769 tab. 3, tab. 4.
- The "nearly 21 mm" claim. The text reports a gain of nearly 21 mm over the prior best 2202.09769 §4.5. In tab. 2, the closest prior entry is PENet at 730.08 mm, and 730.08 − 709.12 = 20.96 mm. The claim is consistent with that entry, but the paper does not name the comparison 2202.09769 tab. 2.
Papers
- Dynamic Spatial Propagation Network for Depth Completion: adaptive affinity per step, diffusion suppression, and three neighbourhood variants (7×7, dilated, deformable) for depth completion.
Source and license
Compiled from the papers listed on this page, all published under CC BY, CC BY-SA or CC0, with citations. See each paper's page for its authors and license.