CEDANet: Citizen-Engaged Domain-Adaptive Network

• CEDANet is a human-in-the-loop weakly supervised domain adaptation framework that segments industrial toxic emissions using full source annotations and citizen-provided video labels.
• It employs a two-stage workflow with pseudo-label generation refined by human guidance and class-aware adversarial feature alignment to address noisy predictions.
• The framework achieves results comparable to limited fully supervised settings by effectively reducing false positives and managing the domain gap in challenging industrial surveillance scenes.

CEDANet, expanded as Citizen-Engaged Domain-Adaptive Network, is a human-in-the-loop weakly supervised domain adaptation framework for industrial toxic emission segmentation. It is designed to segment hazardous industrial smoke or toxic cloud emissions in real surveillance imagery from the IJmond industrial region when dense target-domain pixel annotations are scarce and expensive. The framework addresses two coupled difficulties: industrial smoke segmentation is intrinsically hard because smoke is translucent, amorphous, low-contrast, and easily confused with steam or background haze; and models trained on available source data such as synthetic smoke and wildfire smoke transfer poorly to industrial imagery because of a substantial synthetic/natural-to-industrial domain shift (Tao et al., 29 Jul 2025).

1. Problem setting and conceptual scope

CEDANet is formulated for a setting in which source-domain supervision is available, but target-domain dense masks are largely unavailable. The paper identifies the source-target gap as especially severe because industrial emissions differ from wildfire smoke in color, texture, density, transparency, and scene context, while target images are captured by surveillance cameras under changing weather, lighting, and viewpoints (Tao et al., 29 Jul 2025).

The target-domain annotation bottleneck is central to the method. Fine semantic segmentation already requires extensive manual effort in standard datasets, and the burden is amplified in industrial monitoring because experts must delineate vague smoke boundaries in difficult scenes. CEDANet therefore substitutes dense target supervision with weak supervision from citizens in the form of coarse video-level binary smoke / no-smoke labels. These labels are not used as pixel annotations; instead, they constrain pseudo-label generation and domain adaptation.

At a high level, the framework combines four ingredients: full supervision in the source domain from a labeled smoke segmentation dataset, weak supervision in the target domain from citizen-provided video labels, pseudo-label generation on target videos using a source-trained segmentation model, and class-aware adversarial domain adaptation that aligns smoke and background features separately. The paper’s main claim is that this combination makes it possible to approach limited fully supervised target-domain performance without target-domain pixel annotations (Tao et al., 29 Jul 2025).

2. Two-stage workflow and citizen-guided pseudo-labeling

The CEDANet pipeline is explicitly organized into two stages. In the first stage, a source-supervised segmentation model, denoted $\mathcal{M}_{pretrain}$, is trained on the source dataset and then applied to unlabeled target-domain videos. For each target frame $f$, the model produces a smoke probability map $P(f)\in[0,1]^{H\times W}$ and an initial pseudo-mask. Because these pseudo-labels are noisy under domain shift, the method refines them with a human-guided filtering and temporal selection mechanism (Tao et al., 29 Jul 2025).

Frame confidence is defined by combining mean activation with the fraction of pixels above a smoke threshold:

$$C(f) = \alpha \cdot \mathbb{E}[P(f)] + (1-\alpha)\cdot \frac{\#\{p \in P(f)\mid p>\tau\}}{H\times W}.$$

In implementation, $\alpha=0.85$, and the threshold example is $\tau=0.6$. After all frames in a video are scored, the method selects the top-$k$ candidate center frames rather than only the single best frame. For each candidate, it forms a temporal window with its $\pm n$ neighbors and evaluates temporal coherence using

$$\mathrm{sim}(u,v) = \beta \cdot \mathrm{IoU}(\hat{y}(u),\hat{y}(v)) + (1-\beta)\cdot \mathrm{corr}(P(u),P(v)),$$

where $\hat{y}(u)$ and $f$0 are thresholded binary masks and $f$1 is the Pearson correlation between raw probability maps. The final group score is

$f$2

with $f$3 and $f$4 in implementation. The winning temporal group is then used to generate pseudo-labels. The paper states explicitly that this is not a simple per-frame thresholding procedure, but a video-aware selection process designed to reduce noise and exploit local temporal consistency (Tao et al., 29 Jul 2025).

Citizen supervision enters only at the video level. Each target video receives volunteer-provided coarse labels indicating whether smoke is present. The paper also reports “admin” labels from professional researchers for comparison. The resulting categorical labels and confidence values are as follows:

Category	Value in paper	Confidence
Gold Standard Positive	47	0.9
Strong Positive	23	0.8
Weak Positive	19	0.7
Maybe Positive	5	0.65
Disagreement	3	no constraint
No Data	-1	no constraint

The paper also states that corresponding negative labels cause the system to skip pseudo-labeling or enforce strong negative constraints. It does not specify a more complex probabilistic vote aggregation rule beyond this category-to-confidence mapping.

The weak labels are incorporated by fusing model prediction and a human-derived constraint:

$f$5

The paper then states that $f$6 for positive labels and $f$7 for negative labels. It also notes a slight notation inconsistency because the equation uses both $f$8 and $f$9. The intended meaning is described narratively: positive citizen labels allow model-dominant fusion, while negative labels are more constraint-dominant so that negative human judgments suppress false positives more strongly. If a video is marked disagreement or no data, there is no human constraint and the frame-level pseudo-label is determined solely by the model (Tao et al., 29 Jul 2025).

3. Segmentation backbone, uncertainty modeling, and adaptation objective

CEDANet does not introduce a new segmentation backbone. Its feature generator is built on the Transmission-guided Bayesian (TGB) generative network of Yan et al., which the paper describes as a VAE-based segmentation model with transmission-guided priors and uncertainty calibration. This choice is motivated by the uncertainty and low-contrast boundaries inherent in smoke segmentation (Tao et al., 29 Jul 2025).

The base loss inherited from TGB is

$P(f)\in[0,1]^{H\times W}$0

The exact formulas for $P(f)\in[0,1]^{H\times W}$1, $P(f)\in[0,1]^{H\times W}$2, and $P(f)\in[0,1]^{H\times W}$3 are not restated in the paper. What is specified is the latent-variable mechanism: for an input image $P(f)\in[0,1]^{H\times W}$4, the encoder predicts $P(f)\in[0,1]^{H\times W}$5 and $P(f)\in[0,1]^{H\times W}$6, and the latent code is sampled with the reparameterization trick,

$P(f)\in[0,1]^{H\times W}$7

The sampled latent vector is tiled spatially and concatenated with the input before intermediate processing in a ResNet backbone. The domain discriminators receive features from the third layer of the saliency / segmentation feature generator, so adaptation is performed at an intermediate feature level rather than only at the output layer (Tao et al., 29 Jul 2025).

Feature alignment is class-aware. Instead of using a single global discriminator, CEDANet uses separate discriminators for smoke and background. With feature map $P(f)\in[0,1]^{H\times W}$8 and class mask $P(f)\in[0,1]^{H\times W}$9, masks are formed from thresholded sigmoid outputs:

$$C(f) = \alpha \cdot \mathbb{E}[P(f)] + (1-\alpha)\cdot \frac{\#\{p \in P(f)\mid p>\tau\}}{H\times W}.$$0

The paper implies that the background branch uses the complementary mask. The discriminators $$C(f) = \alpha \cdot \mathbb{E}[P(f)] + (1-\alpha)\cdot \frac{\#\{p \in P(f)\mid p>\tau\}}{H\times W}.$$1 and $$C(f) = \alpha \cdot \mathbb{E}[P(f)] + (1-\alpha)\cdot \frac{\#\{p \in P(f)\mid p>\tau\}}{H\times W}.$$2 receive masked class-specific features.

The discriminators further use attention-guided pooling. Given masked feature map $$C(f) = \alpha \cdot \mathbb{E}[P(f)] + (1-\alpha)\cdot \frac{\#\{p \in P(f)\mid p>\tau\}}{H\times W}.$$3 and class attention map $$C(f) = \alpha \cdot \mathbb{E}[P(f)] + (1-\alpha)\cdot \frac{\#\{p \in P(f)\mid p>\tau\}}{H\times W}.$$4, the pooled feature channel is

$$C(f) = \alpha \cdot \mathbb{E}[P(f)] + (1-\alpha)\cdot \frac{\#\{p \in P(f)\mid p>\tau\}}{H\times W}.$$5

The discriminator design replaces standard convolutions with LDConv (Learnable Deformable Convolution), followed by SiLU activation and global pooling before domain prediction (Tao et al., 29 Jul 2025).

Adversarial alignment is implemented with a Gradient Reversal Layer (GRL) $$C(f) = \alpha \cdot \mathbb{E}[P(f)] + (1-\alpha)\cdot \frac{\#\{p \in P(f)\mid p>\tau\}}{H\times W}.$$6 between the feature generator and the discriminators. In the forward pass, $$C(f) = \alpha \cdot \mathbb{E}[P(f)] + (1-\alpha)\cdot \frac{\#\{p \in P(f)\mid p>\tau\}}{H\times W}.$$7. In backpropagation,

$$C(f) = \alpha \cdot \mathbb{E}[P(f)] + (1-\alpha)\cdot \frac{\#\{p \in P(f)\mid p>\tau\}}{H\times W}.$$8

The paper reports that $$C(f) = \alpha \cdot \mathbb{E}[P(f)] + (1-\alpha)\cdot \frac{\#\{p \in P(f)\mid p>\tau\}}{H\times W}.$$9 is gradually increased from $\alpha=0.85$0 to $\alpha=0.85$1, and that the domain adaptation loss weight is set to $\alpha=0.85$2.

The class-aware domain loss is written as

$\alpha=0.85$3

with

$\alpha=0.85$4

The practical discriminator objective is also written in cross-entropy form as

$\alpha=0.85$5

CEDANet additionally incorporates a local contrastive loss adapted to the binary smoke/background setting. For smoke pixels, the paper gives

$\alpha=0.85$6

with smoke prototype

$\alpha=0.85$7

and per-pixel term

$\alpha=0.85$8

where cosine similarity is

$\alpha=0.85$9

The total objective is stated as

$\tau=0.6$0

The paper also gives an algorithmic notation in which source and target segmentation losses are combined as

$\tau=0.6$1

followed by

$\tau=0.6$2

The text does not clearly reconcile this notation with the earlier total-loss equation containing $\tau=0.6$3, so the exact final implemented loss composition is somewhat under-specified. The paper also does not specify a formal training-time consistency loss; temporal consistency appears only in pseudo-label selection via the group similarity score (Tao et al., 29 Jul 2025).

4. Datasets, annotation regime, and evaluation protocol

The experiments use one source dataset and two target-domain resources. The source dataset is SMOKE5K, containing 5,000 pixel-annotated images: 4,000 synthetic smoke images from SYN70K, 960 real wildfire smoke images, and 40 real smoke images from the Internet. On the target side, the IJmond Camera video dataset contains 878 one-second videos used for pseudo-labeling and adaptation training, and IJmond900 contains 900 fully annotated images from industrial surveillance cameras in the IJmond region (Tao et al., 29 Jul 2025).

The role of IJmond900 is twofold. The paper states that 100 images are used as a gold-standard training set in the limited fully supervised comparison, while the remaining 800 images are used for testing and benchmarking. It also notes that one table contains a typo saying “100 pixel-annotated images for testing only,” but the surrounding text consistently indicates an 800-image test remainder. IJmond900 includes opacity distinctions, with high-opacity and low-opacity smoke regions marked separately.

Preprocessing and evaluation are standardized. Images are cropped into $\tau=0.6$4 patches for evaluation consistency, and all training images are rescaled to $\tau=0.6$5. The reported metrics are Recall, $\tau=0.6$6-score, smoke-class IoU, mean IoU, and mMSE, with additional opacity-specific evaluations (Tao et al., 29 Jul 2025).

A limited fully supervised target-domain comparison is part of the protocol. This comparison uses 100 annotated IJmond900 images and provides an explicit reference point for assessing whether weak citizen feedback plus domain adaptation can approach small-sample full supervision. The paper’s headline framing is that CEDANet reaches a performance regime comparable to this 100-image setting without target-domain annotations for training.

5. Quantitative performance and ablation findings

The principal comparison is reported on $\tau=0.6$7. The paper distinguishes a baseline weak model, a domain-adaptive model without human constraint, CEDANet with citizen feedback, domain adaptation with expert video-level constraints, and a limited fully supervised target-domain setting (Tao et al., 29 Jul 2025).

Setting	$\tau=0.6$8	$\tau=0.6$9
Baseline weak model $k$0	0.083	0.043
Domain adaptation without human constraint $k$1	0.387	0.240
CEDANet with citizen feedback $k$2	0.414	0.261
Expert video-level constraints $k$3	0.448	0.288
Limited fully supervised target setting $k$4	0.418	0.264

The full metric profiles are also reported. The baseline weak model $k$5 has Recall $k$6, mIoU $k$7, and mMSE $k$8. Domain adaptation without human constraint $k$9 has Recall $\pm n$0, mIoU $\pm n$1, and mMSE $\pm n$2. CEDANet with citizen feedback $\pm n$3 has Recall $\pm n$4, mIoU $\pm n$5, and mMSE $\pm n$6. Expert constraints $\pm n$7 yield Recall $\pm n$8, mIoU $\pm n$9, and mMSE $$\mathrm{sim}(u,v) = \beta \cdot \mathrm{IoU}(\hat{y}(u),\hat{y}(v)) + (1-\beta)\cdot \mathrm{corr}(P(u),P(v)),$$0. The limited fully supervised target setting $$\mathrm{sim}(u,v) = \beta \cdot \mathrm{IoU}(\hat{y}(u),\hat{y}(v)) + (1-\beta)\cdot \mathrm{corr}(P(u),P(v)),$$1 yields Recall $$\mathrm{sim}(u,v) = \beta \cdot \mathrm{IoU}(\hat{y}(u),\hat{y}(v)) + (1-\beta)\cdot \mathrm{corr}(P(u),P(v)),$$2, mIoU $$\mathrm{sim}(u,v) = \beta \cdot \mathrm{IoU}(\hat{y}(u),\hat{y}(v)) + (1-\beta)\cdot \mathrm{corr}(P(u),P(v)),$$3, and mMSE $$\mathrm{sim}(u,v) = \beta \cdot \mathrm{IoU}(\hat{y}(u),\hat{y}(v)) + (1-\beta)\cdot \mathrm{corr}(P(u),P(v)),$$4 (Tao et al., 29 Jul 2025).

The headline comparison is that CEDANet with citizen feedback reaches $$\mathrm{sim}(u,v) = \beta \cdot \mathrm{IoU}(\hat{y}(u),\hat{y}(v)) + (1-\beta)\cdot \mathrm{corr}(P(u),P(v)),$$5 and $$\mathrm{sim}(u,v) = \beta \cdot \mathrm{IoU}(\hat{y}(u),\hat{y}(v)) + (1-\beta)\cdot \mathrm{corr}(P(u),P(v)),$$6, versus the baseline’s $$\mathrm{sim}(u,v) = \beta \cdot \mathrm{IoU}(\hat{y}(u),\hat{y}(v)) + (1-\beta)\cdot \mathrm{corr}(P(u),P(v)),$$7 and $$\mathrm{sim}(u,v) = \beta \cdot \mathrm{IoU}(\hat{y}(u),\hat{y}(v)) + (1-\beta)\cdot \mathrm{corr}(P(u),P(v)),$$8. The paper describes this as about a five-fold increase in $$\mathrm{sim}(u,v) = \beta \cdot \mathrm{IoU}(\hat{y}(u),\hat{y}(v)) + (1-\beta)\cdot \mathrm{corr}(P(u),P(v)),$$9 and a six-fold increase in smoke-class IoU. It also emphasizes that citizen-constrained CEDANet is nearly identical to the model trained using 100 fully annotated target images, which scored $\hat{y}(u)$0 in $\hat{y}(u)$1 and $\hat{y}(u)$2 in smoke IoU.

Opacity-specific evaluation shows that all models perform better on high-opacity smoke than on low-opacity smoke. For citizen-constrained CEDANet $\hat{y}(u)$3, the paper reports high-opacity Recall/$\hat{y}(u)$4/IoU of $\hat{y}(u)$5 and low-opacity Recall/$\hat{y}(u)$6/IoU of $\hat{y}(u)$7. Expert constraints improve these values further. The paper interprets this as evidence that faint smoke remains a difficult regime even when the framework improves performance.

The ablation study isolates attention pooling, LDConv, and human feedback constraints. Under citizen constraints, baseline class-aware DA without attention or LDConv gives Recall $\hat{y}(u)$8, $\hat{y}(u)$9, and $f$00. Adding attention pooling only gives Recall $f$01, $f$02, and $f$03. Adding LDConv only gives Recall $f$04, $f$05, and $f$06. The full model with attention + LDConv + citizen constraints gives Recall $f$07, $f$08, and $f$09. Replacing citizen with expert constraints in the full model gives Recall $f$10, $f$11, and $f$12 (Tao et al., 29 Jul 2025).

The paper states that LDConv contributes the biggest recall gain, attention pooling alone is less effective, and human feedback consistently improves pseudo-label quality. It also notes that the temporal pseudo-label refinement heuristic is not separately quantified by an isolated ablation.

6. Interpretation, limitations, and nomenclature

The paper’s interpretation is that CEDANet improves performance by addressing two failure modes simultaneously. First, source-trained segmentation alone fails on industrial imagery because of domain gap. Second, naïve pseudo-labels in the target domain are noisy, especially under severe shift and in the presence of steam. The class-aware adversarial mechanism reduces mismatch at the feature level, while citizen labels remove many obviously wrong pseudo-labels and constrain adaptation through improved target supervision (Tao et al., 29 Jul 2025).

The resulting error profile differs markedly from the weak baseline. The baseline has very high recall but extremely poor $f$13 and IoU, which the paper interprets as heavy smoke overprediction. CEDANet reduces false positives while preserving meaningful smoke detection. Qualitatively, it is described as producing more accurate masks especially in high-opacity regions and achieving a better precision-recall tradeoff.

Several limitations are explicit. A recurring failure case is steam confusion: steam is temporally and spatially coherent and can resemble smoke strongly enough that the model still misclassifies it, especially in positively labeled videos with extensive steam. The paper also stresses that absolute numbers remain moderate because the domain gap is severe and industrial toxic emission segmentation is much harder than standard segmentation benchmarks. On the supervision side, the raw aggregation rule for citizen votes is not specified beyond the category-confidence mapping. On the optimization side, the exact implemented loss composition is somewhat under-specified, and no formal consistency loss is defined beyond temporal coherence in pseudo-label selection (Tao et al., 29 Jul 2025).

The practical significance claimed for the framework is its scalability and cost-efficiency. The paper presents CEDANet as a way to convert cheap binary video judgments from citizens into useful dense supervision for environmental monitoring, thereby avoiding the cost of large target-domain mask datasets. A plausible implication is that the framework is meant for long-term surveillance scenarios in which new weakly labeled videos can continue to feed pseudo-labeling and adaptation.

CEDANet should also be distinguished from CADENet, the separate adverse-weather perception method introduced in “CADENet: Condition-Adaptive Asynchronous Dual-Stream Enhancement Network for Adverse Weather Perception in Autonomous Driving” (Khairy et al., 19 May 2026). The similarity in name can cause confusion, but the two systems address different tasks, use different architectures, and are presented in different papers. CEDANet refers specifically to the citizen-engaged domain-adaptive framework for industrial toxic emission segmentation (Tao et al., 29 Jul 2025).