S23DR 2026 Challenge Overview

• S23DR 2026 Challenge is a multi-task benchmark featuring domain-agnostic incremental learning for sound classification, quality-aware infrared super-resolution, and structured 3D wireframe reconstruction.
• Evaluation metrics such as domain-balanced accuracy, PSNR+20×SSIM, and Hybrid Structure Score emphasize structural fidelity and operational robustness across heterogeneous tasks.
• Methods leveraging data augmentation, domain-specific adaptations, and ensemble consensus are crucial for overcoming challenges posed by non-ideal inputs and incremental learning protocols.

The S23DR 2026 Challenge denotes a 2026 challenge umbrella documented through multiple task reports and a winning-solution paper rather than through a single homogeneous benchmark definition. The available records associate S23DR 2026 with at least three distinct problem settings: Task 7 of the DCASE 2026 Challenge / S23DR 2026 on domain-agnostic incremental learning for sound classification, the NTIRE 2026 / S23DR 2026 Remote Sensing Infrared Image Super-Resolution (×4) Challenge, and a challenge on structured 3D wireframe reconstruction from sparse SfM, fitted depth, and semantic segmentations (Casciotti et al., 1 Jun 2026, Liu et al., 23 Apr 2026, Skvrna et al., 4 Jun 2026). In all three cases, the benchmarked systems are evaluated under conditions that depart from idealized assumptions: sequential domain exposure without replay and without domain labels at inference, single-channel infrared reconstruction under heterogeneous image quality, and non-RGB 3D structure recovery from sparse and incomplete geometric evidence.

1. Scope and task portfolio

Published S23DR 2026-associated papers describe a portfolio of technically heterogeneous tasks spanning audio classification, remote sensing image restoration, and structured 3D scene reconstruction. A concise summary of the documented tasks is given below (Casciotti et al., 1 Jun 2026, Liu et al., 23 Apr 2026, Skvrna et al., 4 Jun 2026).

Task	Core problem	Official evaluation
DCASE 2026 Task 7 / S23DR 2026	Domain-agnostic incremental learning for sound classification across $D_1,D_2,D_3$ with 10 fixed sound classes	Overall accuracy averaged over the three domains
NTIRE 2026 / S23DR 2026 Infrared SR	Single-image $4\times$ super-resolution for single-channel infrared remote sensing images	$\text{PSNR} + 20 \times \text{SSIM}$
Structured 3D wireframe reconstruction	Predict a labeled 3D graph from sparse COLMAP SfM, fitted depth maps, and semantic segmentations	Hybrid Structure Score (HSS)

This task diversity is a defining characteristic. The audio task formalizes domain-incremental learning in a domain-agnostic setting; the infrared task formalizes single-track infrared super-resolution with a metric that heavily weights structural fidelity; and the 3D task formalizes structured graph prediction from sparse geometric and semantic cues. A plausible implication is that S23DR 2026 functioned less as a single benchmark than as a challenge program emphasizing robustness, structure, and deployment-relevant failure modes across modalities.

2. Domain-agnostic incremental learning for sound classification

In Task 7 of the DCASE 2026 Challenge / S23DR 2026, the objective is to learn 10 sound classes—alarm, baby_cry, bark, engine, fire, footsteps, knock, telephone_ringing, piano, speech—across three domains $D_1, D_2, D_3$ under a strict incremental protocol (Casciotti et al., 1 Jun 2026). The model first learns from the first domain, then from the next, then from the third, and at each incremental step it has access only to the current domain’s data. After learning $D_2$, it does not get to replay $D_1$; after learning $D_3$, it does not get to revisit $D_1$ or $D_2$. At inference time, the model receives only the audio sample, no domain label is given, and it must classify the sample into one of the 10 target classes.

The task is explicitly framed as domain-incremental learning (DIL) rather than class-incremental learning because the classes are fixed and the domain changes. It is also distinguished from domain-aware DIL because domain identity is unavailable at test time. The paper argues that this makes the problem more practical than standard two-domain adaptation and more difficult than domain-aware continual learning, since the model must classify samples from all seen domains without explicit domain identity.

The development dataset contains data from D2 and D3, with D1 knowledge embedded in the baseline system. The development data include 139 minutes from D2 and 275 minutes from D3. The evaluation dataset contains files from all three domains, but only audio is provided and no domain information is given. Each clip has one target label, although background sounds may also be present.

The task imposes stringent constraints. External data is not allowed. Pretrained models are not allowed. Only data provided in the task may be used. Allowed operations include data augmentation, mixing data sampled from a distribution, pitch shifting, and time stretching. Participants may not annotate the evaluation set, the evaluation set cannot be used for training, no statistics from the evaluation set may be used in decision-making, and classification must be done independently for each test sample.

Evaluation is based on domain-balanced and class-balanced accuracy. Accuracy is calculated separately for each domain, each domain accuracy is the average of the class-wise accuracies, and the final score is the average over the three domains. The baseline system is a convolutional neural network with 6 convolutional blocks, 2 convolutional layers + batch normalization per block, global pooling, a fixed-length embedding, and a softmax classifier, following the structure of PANNs CNN14. It is trained for 120 epochs with batch size 32, Adam, a learning rate of 0.0001 in the initial phase and 0.00001 in incremental phases. Audio is resampled to 32 kHz, segmented into 4-second clips, and converted into log mel-band energies with 64 mel bands, a lower cutoff of 50 Hz, an upper cutoff of 14 kHz, a 1024-sample Hamming window, and a 320-sample hop.

Its domain-agnostic incremental mechanism uses domain-specific batch normalization layers. The model is first trained on D1; for new domains $D_2$ and $4\times$0, it keeps domain-shared convolutional layers and adapts separate BN layers for each domain. At inference time, the input is passed through the shared layers combined with each available domain-specific BN set, each path outputs class probabilities, the system computes the entropy of the predicted distribution, the BN path with minimum entropy is selected, and its output is used for the final classification. The reported baseline achieves 54.7% accuracy on D2 after learning it, 35.0% on D3 after learning it, and maintains 54.7% on D2 after learning D3, yielding 44.9% average accuracy over the two available development domains. With the correct BN layers selected for each test item, the comparison point rises to 71.8% on D2, 63.4% on D3, and 67.6% on average, indicating that erroneous inference of the test sample domain is the main bottleneck.

3. Remote sensing infrared image super-resolution

The NTIRE 2026 / S23DR 2026 Remote Sensing Infrared Image Super-Resolution Challenge formalizes a single-image super-resolution problem at 4× upscaling for single-channel infrared data (Liu et al., 23 Apr 2026). The low-resolution images are generated from high-quality infrared images by bicubic downsampling with scale factor 4, and the task is to predict a high-resolution infrared image that matches the hidden ground truth. The benchmark is explicitly framed for remote sensing applications in which recovering fine spatial detail matters for target detection, environmental monitoring, and change analysis.

The challenge uses a single-track design. The official evaluation score is

$4\times$1

This weighting gives structural preservation a large role: an SSIM improvement of 0.01 is equivalent to a 0.2 dB PSNR gain under the official score. All metrics are computed on the single infrared channel, even when systems internally replicate grayscale data to three channels in order to reuse RGB-pretrained models.

The official InfraredSR dataset is divided into training, validation, and test subsets. The training set contains 625 images at 320 × 256, 281 images at 120 × 120, 99 images at 64 × 64, 9 images at 256 × 256, and 5 images at 160 × 128. The validation set contains 100 images, and the test set contains 222 images, including 188 images at 1280 × 1024, 22 images at 480 × 480, and 12 images at 256 × 256. The challenge credits Spark Transmission (Beijing) Co., Ltd. for data support.

The evaluation protocol is Codalab-based. In the development phase, participants had access to the full training and validation sets, trained on the training set, and uploaded validation restorations for real-time ranking. In the testing phase, participants received only the LR test images and uploaded SR outputs to the Codalab server while also submitting code and a report. Final ranking was based on organizers’ hidden-test results, with submitted code used for reproducibility checks. External data were allowed only if they did not overlap with validation or test data.

The challenge attracted 115 registered participants, of whom 13 teams submitted valid entries. The top-ranked team, WHU-VIP, achieved 35.9643 PSNR, 0.9236 SSIM, and 54.4361 score. The top five teams—WHU-VIP, XJRes, FengFans, I2WM…JNU, and davinci—were separated by only about 0.15 dB, indicating a tightly clustered leaderboard.

The strongest submissions converged on several architectural themes. WHU-VIP proposed QAHAT, a quality-aware adaptation of HAT (Hybrid Attention Transformer) with a Global Quality Branch, Local Quality Branch, Quality Estimation Module (QEM), Source Quality Adapters (SQA), and Local Quality-Aware Module (LQAM). XJRes combined Progressive Focused Transformer (PFT) and HAT in a four-branch ensemble with weighted averaging. FengFans used HAT-L, two separately trained models, and 8-fold geometric test-time augmentation using the dihedral group $4\times$2, and explicitly identified TTA as its largest single source of test-time gain. The report also identifies recurring design patterns across teams: Transformer-based architectures, state-space models (Mamba), hybrid Transformer-CNN systems, frequency-aware learning, and progressive or curriculum training.

4. Structured 3D wireframe reconstruction

A separate S23DR 2026 task addressed structured 3D wireframe reconstruction from non-RGB inputs, with the winning solution reported in “S23DR 2026 Winning Solution” (Skvrna et al., 4 Jun 2026). The goal is to predict a labeled 3D graph

$4\times$3

where $4\times$4 are metric 3D vertices such as roof apexes, ridge endpoints, eave corners, and building corners, and $4\times$5 are straight edges connecting those vertices. Each edge also has a semantic class determined by adjacent faces, such as rake, eave, flashing, ground line, and related categories. The task input does not provide RGB images. Instead, each scene is represented by a sparse COLMAP SfM point cloud with camera poses and intrinsics, metric depth maps fitted to the SfM scale, and semantic segmentations projected into each view.

Predictions are scored using the Hybrid Structure Score (HSS), which combines a vertex term—F1 of predicted versus ground-truth vertices under optimal Hungarian assignment with a matching radius—and an edge term—volumetric IoU after rendering edges as cylinders and intersecting predicted and ground-truth edge solids. The score is the harmonic mean of these two terms, so both geometric localization and graph topology matter.

The winning method treats reconstruction as conditional set generation in 3D. It denoises 64 vertex tokens with a flow-matching DiT conditioned on Perceiver-style scene tokens. The pipeline is:

$4\times$6

The scene encoder uses enriched point representations including normalized 3D coordinates, Fourier encoding of position, a point-type embedding for COLMAP, depth-derived point, or camera center, Gestalt semantic labels, an ADE20K class embedding, and confidence scalars. Coordinates are centered by the median COLMAP location and scaled by a robust 95th-percentile radius. Stage 1 uses a fixed budget of 8192 points, with half reserved for COLMAP and half for depth points, while Stage 2 uses 16384 points inside the crop.

The Perceiver-style encoder contextualizes the points, introduces 1024 learned pooling tokens anchored at real scene points selected by priority-aware Gestalt farthest-point sampling, and produces a fixed scene representation of 1024 scene tokens. The output wireframe is represented as a set of 64 vertex slots $4\times$7. During training, ground-truth vertices are assigned to slots using Hungarian matching. The DiT blocks cross-attend to the time embedding and the scene tokens, and for each vertex slot the model predicts a velocity in $4\times$8, a validity logit, and pairwise edge logits. The total loss is

$4\times$9

At inference, the learned ODE is integrated using 50 Euler steps. A two-pass inference strategy first predicts a coarse global structure and then refines it in a hull-cropped region. To stabilize stochastic sampling, the system runs 16 stochastic trajectories, selects the medoid prediction using a distance based on Hungarian-aligned vertex distance and edge agreement, and replaces each selected vertex with a consensus aggregate computed from matched vertices across the other samples.

The reported outcome is 1st place on the private leaderboard with HSS = 0.654. The paper also reports that second place achieved 0.648, the organiser learned baseline 0.474, the organiser handcrafted baseline 0.391, and the winning system had the highest vertex F1 among submissions: 0.791. On validation, the reported progression is 0.4764 for Stage 1 large, 0.4977 for Stage 1 + Stage 2, and 0.5025 for the 16-sample ensemble.

5. Evaluation logic and benchmark design

Although the documented S23DR 2026 tasks are heterogeneous, their evaluation protocols share a common emphasis on operational robustness rather than isolated pointwise accuracy. In the sound-classification task, the score is the average of domain-wise accuracies, and each domain accuracy is itself the average of class-wise accuracies, ensuring both domain balance and class balance (Casciotti et al., 1 Jun 2026). In the infrared super-resolution task, the official score weights SSIM heavily relative to PSNR, thereby favoring structural consistency rather than pure pixel fidelity (Liu et al., 23 Apr 2026). In the wireframe reconstruction task, HSS is the harmonic mean of a vertex-detection term and an edge-overlap term, penalizing systems that recover only geometry or only topology (Skvrna et al., 4 Jun 2026).

The benchmark constraints are similarly task-shaped. The audio task forbids external data and pretrained models, forcing progress through incremental-learning design rather than scale. The infrared task allows external data only under a non-overlap condition, and official evaluation is always performed on the single infrared channel even when models internally use RGB-style processing. The wireframe task is notable for its non-RGB input regime: sparse SfM, fitted depth, and semantic projections replace conventional image-based reconstruction inputs.

This design also clarifies a possible misconception. S23DR 2026 is not documented, in the available records, as a single leaderboard with a unified metric or a single modality. Instead, the documented tasks instantiate different notions of robustness: catastrophic forgetting and hidden domain identity in audio, heterogeneous degradation and structural fidelity in infrared super-resolution, and metric graph recovery from sparse geometric evidence in 3D reconstruction.

6. Methodological patterns and research significance

Across its documented tasks, S23DR 2026 rewarded methods that explicitly encode nuisance variability rather than assuming it away. In the sound-classification task, the baseline retains domain-shared convolutional layers while adapting separate domain-specific BN layers, and its primary failure mode is not class confusion alone but domain misidentification inside the model’s routing mechanism (Casciotti et al., 1 Jun 2026). In the infrared super-resolution task, the leading methods model quality heterogeneity, use transformer-based backbones, and rely heavily on ensemble fusion and test-time augmentation (Liu et al., 23 Apr 2026). In the wireframe task, the winning solution uses a Perceiver-style scene encoder, conditional set generation, flow matching, and multi-sample consensus to manage sparse, noisy, and incomplete inputs (Skvrna et al., 4 Jun 2026).

This suggests a broader interpretation of S23DR 2026 as a challenge program centered on structured robustness. The audio task emphasizes learning across domains without replay and without explicit domain identity at inference. The infrared task emphasizes reconstructing structurally consistent and thermally faithful outputs from bicubic $\text{PSNR} + 20 \times \text{SSIM}$0 downsampling. The 3D task emphasizes metrically accurate graph prediction from sparse geometry and semantics rather than from dense RGB appearance. In each case, high performance depends on architectures that preserve structure under shift: domain-balanced classifiers, quality-aware restoration networks, and scene-conditioned generative set models.

The documented outcomes also indicate that strong baselines remain substantially below the best achievable performance under these conditions. The audio baseline reaches 44.9% average accuracy over the two available development domains and is limited mainly by incorrect BN-path selection. The infrared leaderboard is tightly packed, showing that small design choices around structural modeling and inference can alter rank. The 3D winning solution improves markedly over the organiser learned baseline 0.474 and handcrafted baseline 0.391, while still leaving headroom above HSS = 0.654. Taken together, these results position S23DR 2026 as a set of benchmarks in which robustness is operationalized through concrete protocol design rather than treated as a secondary diagnostic property.