CertainNet: Uncertainty-Aware Graph Detector

• CertainNet is an uncertainty-aware, graph-based detection framework that refines object proposals with high epistemic uncertainty, improving recall and precision in aerial images.
• It employs MC-dropout to quantify uncertainty and partitions proposals into 'certain' and 'uncertain' sets, allowing targeted refinement with minimal extra computational cost.
• The method uses spatial and semantic edge construction for graph creation, enabling directed message passing from confident detections to ambiguous ones, achieving notable mAP gains on DOTA benchmarks.

The Uncertainty-Aware Graph network for object DETection (UAGDet) is a detection framework designed for aerial images with densely packed and variably sized objects. UAGDet integrates uncertainty estimation into a graph-based refinement process, selectively contextualizing and improving detections for proposal regions with high epistemic uncertainty. Its architecture extends general two-stage detectors (e.g., Faster R-CNN, RoITransformer) and introduces a graph neural network (GNN) that propagates context from “certain” to “uncertain” objects. This yields substantial gains on benchmarks such as DOTA, especially for ambiguous detections in crowded or cluttered scenes (Kim et al., 2022).

1. Conceptual Foundations

UAGDet addresses limitations in standard CNN-based two-stage object detectors, particularly their tendency to underperform on ambiguous or small objects in aerial images. The method’s core insight is to identify “uncertain” proposals using Monte Carlo dropout-based uncertainty measurement, and to refine only these high-uncertainty regions via directed message passing in a spatial-semantic object graph. By limiting computational cost and overfitting risk, UAGDet circumvents inefficiencies and redundancy associated with refining all object proposals equally.

Key design elements include:

• Uncertainty quantification via lightweight MC-dropout.
• Proposal partition into “certain” (low uncertainty) and “uncertain” (high uncertainty) sets.
• Graph construction linking certain nodes to uncertain ones based on spatial and semantic similarity.
• GCN-based feature refinement applied solely to uncertain detections.
• Uncertainty-weighted loss to scale backpropagation focus.

2. Detection and Uncertainty Estimation

The base detection pipeline leverages an RPN and RoIAlign to extract proposal-level features $\hat{f}_i^{\mathrm{conv}}$ for each region. The detector’s classification ($\widehat{p}_i$) and regression ($\widehat{\mathrm{bbox}}_i$) heads process these for initial predictions. Dropout is applied at the final fully connected layers, with MC-dropout used for epistemic uncertainty estimation:

$\phi_i = \mathrm{Stdev}(\{p_i^{(m)}\}_{m=1}^M)$

where $p_i^{(m)}$ is the classification head output for proposal $i$ under the $m$-th MC-dropout trial ($M=50$ in practice). All proposals are subjected to a minimal NMS, and the retained set is partitioned by $\phi_i$, the standard deviation of predicted logits.

3. Graph Construction and Message Passing

A directed bipartite graph $\mathcal{G}$ is constructed over the retained proposals, partitioned as:

$\widehat{p}_i$0

Edges $\widehat{p}_i$1 are included if $\widehat{p}_i$2, $\widehat{p}_i$3, and both

$\widehat{p}_i$4

where spatial distance is

$\widehat{p}_i$5

and semantic distance operates on feature vectors

$\widehat{p}_i$6

Edge weights

$\widehat{p}_i$7

encode the inverse normalized spatial proximity.

Node features are composed via 1×1 convolutional reduction on RoI features concatenated with embedded class predictions:

$\widehat{p}_i$8

A two-layer GCN updates only the uncertain node features by aggregating messages from connected certain nodes. The GCN output per uncertain node is concatenated with the original CNN features, forming the input to a refined classification head.

4. Loss Function and Training Objective

Uncertainty-aware supervision is implemented by scaling losses according to the normalized uncertainties $\widehat{p}_i$9. Each $\widehat{\mathrm{bbox}}_i$0 is processed through a softmax with temperature $\widehat{\mathrm{bbox}}_i$1:

$\widehat{\mathrm{bbox}}_i$2

The refined classification loss is:

$\widehat{\mathrm{bbox}}_i$3

The total training objective aggregates the standard and refined losses:

$\widehat{\mathrm{bbox}}_i$4

with $\widehat{\mathrm{bbox}}_i$5 in reported experiments.

This structure directs optimization pressure primarily onto hard cases, producing models with improved recall and precision for uncertain object proposals.

5. Algorithmic Workflow

The principal steps of UAGDet are as follows:

1. Initial Proposals: Generate region proposals and extract RoI-aligned features from the backbone CNN.
2. Uncertainty Estimation: Apply MC-dropout at the classification/regression heads to estimate $\widehat{\mathrm{bbox}}_i$6 for each proposal via variance over predictions.
3. Node Partitioning: Retain post-NMS objects, ranking by $\widehat{\mathrm{bbox}}_i$7 and splitting into $\widehat{\mathrm{bbox}}_i$8 (certain) and $\widehat{\mathrm{bbox}}_i$9 (uncertain).
4. Graph Formation: For each $\phi_i = \mathrm{Stdev}(\{p_i^{(m)}\}_{m=1}^M)$0 and $\phi_i = \mathrm{Stdev}(\{p_i^{(m)}\}_{m=1}^M)$1, create edges when spatial and semantic conditions are met, assign $\phi_i = \mathrm{Stdev}(\{p_i^{(m)}\}_{m=1}^M)$2.
5. GCN Refinement: Build per-node features, stack into $\phi_i = \mathrm{Stdev}(\{p_i^{(m)}\}_{m=1}^M)$3, and perform two-layer GCN propagation, updating only uncertain nodes.
6. Prediction and Loss: For each uncertain node, concatenate GCN features, apply the refined classification head, and compute the uncertainty-weighted loss for training.

At inference, the same procedure is followed with dropout inactive.

6. Experimental Evaluation

UAGDet was evaluated on the DOTA-v1.0 and DOTA-v1.5 datasets (up to 403,318 objects over 2,806 high-resolution images). The base detector is RoITransformer with ResNet-101 FPN backbone, using 1,250 initial proposals.

The following table summarizes benchmark results:

Detector	DOTA-v1.0 mAP	DOTA-v1.5 mAP
Mask R-CNN	70.7	62.7
GFNet	71.6	-
RoITransformer	73.4	65.5
UAGDet	75.1	68.0

UAGDet improves mAP by +1.7 and +2.5 over RoITransformer on DOTA-v1.0 and v1.5, respectively. Ablation studies attribute 0.4–1.2 mAP improvement each to spatial-semantic edge construction, uncertainty-aware GCN message passing, and loss weighting.

7. Significance and Implications

UAGDet demonstrates significant accuracy improvements in dense-object aerial image detection via targeted uncertainty-driven message passing. Focusing computational and learning resources on ambiguous regions allows models to better handle the core challenges of spatial clutter and ambiguous semantic classes, common in satellite and surveillance imagery. The architecture is modular, allowing insertion into any RoI-based two-stage detector pipeline with only minor computational overhead.

The methodology suggests broader applicability of hybrid uncertainty-graph refinement in other perception domains characterized by densely distributed ambiguous objects, with strong empirical gains for small and hard instances (Kim et al., 2022).