ACDC: Dataset for Adverse Conditions

• ACDC is a large-scale dataset offering adverse-condition driving scenes with paired normal-condition references and uncertainty-aware annotations.
• The dataset enables robust semantic segmentation, object detection, instance and panoptic segmentation, and domain adaptation via high-quality pixel-level labeling and explicit invalid region modeling.
• Its design supports diverse tasks including cross-domain adaptation and uncertainty-aware predictions, establishing a benchmark for robust perception under challenging weather.

The Adverse Conditions Dataset with Correspondences (ACDC) is a large-scale benchmark for semantic driving scene perception under adverse visual conditions, introduced to address the dominance of daytime, clear-weather imagery in prior autonomous-driving datasets and the consequent fragility of perception systems outside nominal conditions (Sakaridis et al., 2021). It consists of 8012 images, including 4006 adverse-condition images equally distributed between fog, nighttime, rain, and snow, together with corresponding normal-condition images of the same scenes and privileged annotations designed to expose both semantic content and irreducible ambiguity under degraded visibility (Sakaridis et al., 2021). ACDC was designed for robust semantic segmentation, object detection, instance segmentation, panoptic segmentation, and uncertainty-aware semantic segmentation, and it has become a central benchmark for supervised learning, unsupervised domain adaptation, source-free adaptation, continual adaptation, and augmentation-based robustness studies in adverse-weather perception (Sakaridis et al., 2021, Bruggemann et al., 2022, Bruggemann et al., 2023, Lee et al., 2024, Wang et al., 25 Mar 2026).

1. Motivation and problem setting

ACDC was introduced from the premise that level-5 driving automation requires a robust visual perception system that can parse input images under any condition, whereas existing driving datasets for dense semantic perception are either dominated by images captured under normal conditions or are small in scale (Sakaridis et al., 2021). Large-scale semantic-driving datasets such as Cityscapes, Mapillary Vistas, and KITTI have very limited coverage of adverse conditions, while datasets covering adverse weather such as Oxford RobotCar and BDD100K either lack dense pixel-level annotations or provide only limited and often unreliable adverse-condition annotations (Sakaridis et al., 2021).

A second motivation was that existing pixel-level annotations for adverse conditions generally did not account for aleatory uncertainty: some image regions are intrinsically ambiguous because of darkness, sensor noise, fog, snowfall, lens artifacts, or related degradations (Sakaridis et al., 2021). ACDC therefore coupled dense semantic annotation with an explicit notion of invalidity, so that ambiguous regions could be separated from regions whose semantics remain visually ascertainable (Sakaridis et al., 2021). This design choice also underpins the dataset’s uncertainty-aware segmentation task.

A third motivation was methodological rather than purely curatorial. ACDC was constructed with corresponding normal-condition images of the same scenes, enabling adaptation and cross-condition transfer methods to exploit weak supervision from easier views of the same locations (Sakaridis et al., 2021). This property later became central to alignment- and contrast-based adaptation methods such as Refign and CMA, which explicitly use the reference-target pairing provided by ACDC (Bruggemann et al., 2022, Bruggemann et al., 2023).

2. Composition, acquisition, and scene taxonomy

ACDC consists of a large set of 8012 images, half of which, 4006, are equally distributed between four common adverse conditions: fog, nighttime, rain, and snow (Sakaridis et al., 2021). The adverse subset contains 1000 fog images, 1006 nighttime images, 1000 rain images, and 1000 snow images (Sakaridis et al., 2021). Each adverse-condition image comes with a corresponding image of the same scene under normal conditions, and 1503 of the corresponding normal-condition images feature panoptic annotations, raising the total number of annotated images to 5509 (Sakaridis et al., 2021).

The dataset uses the same 19 semantic classes as Cityscapes, enabling direct transfer and comparison in semantic segmentation and domain adaptation settings (Sakaridis et al., 2021, Li et al., 2022). For instance- and panoptic-level tasks, ACDC provides annotations for “things” classes such as vehicles and humans, alongside “stuff” classes in standard panoptic form (Sakaridis et al., 2021).

The following summary organizes the core dataset statistics reported for ACDC.

Component	Reported content
Total images	8012
Adverse images	4006
Adverse conditions	Fog, nighttime, rain, snow
Annotated normal-condition images	1503
Total annotated images	5509
Semantic classes	19, matching Cityscapes

Data recording was carried out with a GoPro Hero 5 at 1080p, 8-bit RGB, and 30 Hz (Sakaridis et al., 2021). The adverse subsets were recorded so that each acquisition emphasizes a single adversity: fog was recorded under foggy daytime, nighttime strictly at night, rain in daylight with clear rain effects, and snow with both falling snow and snow cover on the ground (Sakaridis et al., 2021). This strict visual separation increases domain shift and makes the benchmark deliberately difficult (Sakaridis et al., 2021).

All adverse conditions and their reference images are split into train, validation, and test according to a strict geographic split with zero overlap, and test labels are withheld for benchmarking (Sakaridis et al., 2021). This geographic separation is important because it limits trivial memorization of locations and preserves the benchmark’s role as a cross-condition generalization testbed.

3. Annotation protocol and uncertainty modeling

A defining feature of ACDC is its annotation protocol for adverse imagery. Each adverse-condition image has high-quality pixel-level panoptic annotation, but the annotation is produced in a two-stage process specifically intended to separate visible semantics from ambiguity (Sakaridis et al., 2021).

In stage 1, a human annotator labels what can be unambiguously determined from the adverse image alone (Sakaridis et al., 2021). In stage 2, the annotator cross-references the corresponding normal-condition image and video, revisiting ambiguous regions; if a region can now be assigned a correct label using this privileged information, it is labeled, whereas regions that remain indiscernible stay unlabeled and are excluded (Sakaridis et al., 2021). Each adverse-condition annotation therefore includes a binary mask, denoted $J$, indicating invalid regions: pixels whose label changes between stage 1 and stage 2 are marked invalid, while the remainder are treated as valid (Sakaridis et al., 2021).

This protocol yields a benchmark in which not all pixels are equally observable. Reported annotation statistics indicate that about 85% of pixels in adverse images are labeled as valid, 7.5% are labeled as invalid after the privileged step, and the remainder are unlabelable (Sakaridis et al., 2021). ACDC therefore differs from conventional semantic segmentation datasets in that uncertainty is not treated merely as model error; part of it is embedded in the ground-truth specification itself.

That design supports a distinct task: uncertainty-aware semantic segmentation (Sakaridis et al., 2021). In this setting, models predict both semantic labels and a confidence map per pixel, and evaluation uses Average Uncertainty-aware IoU (AUIoU), which averages UiOU over confidence thresholds while rewarding systems that assign low confidence to truly invalid regions and penalizing overconfident predictions in ambiguous areas (Sakaridis et al., 2021). The benchmark thus formalizes a setting in which abstention can be correct.

A frequent misconception is that ACDC’s “correspondences” imply dense pixel-wise correspondence labels between normal and adverse images. The dataset does not provide pixel- or instance-level correspondences; rather, it provides image-level correspondences of the same scene and approximate viewpoint, obtained by synchronizing GPS logs from adverse and normal recordings with dynamic-programming-based sequence alignment (Sakaridis et al., 2021, Bruggemann et al., 2022). Later methods such as Refign and CMA estimate dense alignments on top of these image-level pairs rather than reading them directly from the dataset (Bruggemann et al., 2022, Bruggemann et al., 2023).

4. Correspondences and supported perception tasks

The “with Correspondences” component of ACDC’s name refers to its paired normal/adverse images of the same scenes (Sakaridis et al., 2021). These correspondences were created by synchronizing GPS logs from adverse and normal recordings, producing approximate viewpoint matches that can serve both annotation and learning (Sakaridis et al., 2021). Their first role is curatorial: privileged normal-condition information improves labeling quality for adverse scenes. Their second role is algorithmic: they provide weak supervision for cross-condition adaptation, feature alignment, and robustness analysis (Sakaridis et al., 2021, Bruggemann et al., 2022, Bruggemann et al., 2023).

ACDC supports five benchmark tasks: semantic segmentation, panoptic segmentation, instance segmentation, object detection, and uncertainty-aware semantic segmentation (Sakaridis et al., 2021). This breadth made it possible for the dataset to function as more than a segmentation benchmark. In later object-detection experiments, ACDC was used as a real-world test set for evaluating robustness under adverse weather, and because it does not directly provide bounding boxes for traffic lights and traffic signs, those labels were generated from provided semantic segmentation masks using the skimage.measure.regionprops Python function (Gurbindo et al., 13 May 2025). In that study, evaluation used 250 images for each adverse weather condition from ACDC, for a total of 1000 ACDC images (Gurbindo et al., 13 May 2025).

The correspondence structure has also been central to semantic adaptation work. Refign treats ACDC reference-target pairs as weak supervision and augments self-training-based unsupervised domain adaptation with uncertainty-aware dense alignment and adaptive label correction (Bruggemann et al., 2022). CMA studies source-free model adaptation under the constraint that source data are inaccessible, using ACDC’s GPS-matched normal/adverse image pairs to learn condition-invariant features via contrastive learning (Bruggemann et al., 2023). FREST likewise operates in a source-free setting on ACDC, alternating condition-embedding learning and feature restoration using unlabeled adverse images paired with reference images under normal conditions from similar geolocations (Lee et al., 2024).

This suggests that ACDC’s principal scientific contribution is not only the adverse imagery itself but the coupling of adverse visual degradation with a weakly aligned, normal-condition companion domain. A plausible implication is that the dataset supports a broader class of adaptation paradigms than benchmarks that provide adverse images alone.

5. ACDC as a benchmark for robust segmentation and adaptation

ACDC rapidly became a standard target domain for semantic segmentation under adverse conditions. In “Semantic Segmentation under Adverse Conditions: A Weather and Nighttime-aware Synthetic Data-based Approach,” ACDC is used only for evaluation, with the validation split of 506 images spanning rainy, foggy, snowy, and nighttime conditions (Kerim et al., 2022). In that study, a DeepLabV3+-based model with weather-aware and time-aware supervisors achieved an overall ACDC mIoU of 0.49, compared with 0.35 for the DeepLabV3+ baseline, while maintaining 0.75 mIoU on Cityscapes (Kerim et al., 2022). The paper reports that baseline models trained on Cityscapes only perform poorly on ACDC, with overall mIoU in the range 0.35–0.42, and that naive fine-tuning on synthetic adverse weather data improves ACDC but can substantially reduce Cityscapes performance (Kerim et al., 2022).

WeatherProof re-examines ACDC from the perspective of language-guided segmentation. Using InternImage and CLIP-derived weather composition “side information,” it reports validation-set results on rain, fog, and snow, with average mIoU increasing from 75.3 for InternImage to 81.0 for InternImage + the proposed method, and an mIoU increase from 76.17 to 82.6 in another reported ACDC setting (Gella et al., 2024). The study emphasizes that although ACDC divides data by primary weather effect, real-world images can contain multiple overlapping degradations, and that language-guided weather composition modeling is particularly effective in such mix scenarios (Gella et al., 2024).

In unsupervised domain adaptation, ACDC has functioned as the canonical adverse-weather target for Cityscapes-to-adverse transfer. VBLC evaluates Cityscapes $\rightarrow$ ACDC on the official test set and reports 47.8 mIoU with a DeepLab-v2 backbone versus 45.7 for FDA, and 64.2 mIoU with a SegFormer backbone versus 55.3 for DAFormer (Li et al., 2022). Heuristic Self-Paced Learning later reports state-of-the-art Cityscapes $\rightarrow$ ACDC performance of 72.9 mIoU on the ACDC test set and 72.7 on the validation set when built on HRDA (Wang et al., 25 Mar 2026).

In source-free adaptation, ACDC’s correspondences are especially valuable because they compensate, in part, for the absence of source data during adaptation. CMA reports 69.1 mIoU on the ACDC test set with a SegFormer backbone, improving over source-free baselines such as TENT, HCL, URMA, and URMA+SimT (Bruggemann et al., 2023). FREST reports 70.7 mIoU on Cityscapes $\rightarrow$ ACDC, improving over CMA’s 69.1, and also reports per-condition gains on fog, night, rain, and snow (Lee et al., 2024).

ACDC has also been used to study continual multi-target adaptation across adverse conditions rather than one-shot adaptation to a single target. In “Semantic Segmentation in Multiple Adverse Weather Conditions with Domain Knowledge Retention,” the four ACDC splits are treated as sequential target domains in the order Cityscapes $\rightarrow$ Night $\rightarrow$ Rain $\rightarrow$ Fog $\rightarrow$ Snow (Yang et al., 2024). With a DAFormer backbone, the proposed method reports an averaged mIoU of 65.7 and accumulated forgetting of 3.6, compared with 60.1 and 11.3 for MIC (Yang et al., 2024). This use of ACDC highlights that the dataset can be decomposed into weather-specific target domains for studying knowledge retention and catastrophic forgetting.

6. Empirical difficulty, error modes, and broader impact

ACDC is widely regarded as exceptionally challenging because adverse conditions induce severe visibility degradation, large domain shifts, and condition-specific artifacts (Gurbindo et al., 13 May 2025, Kerim et al., 2022). Reported causes include rain droplets, snow flakes, fog acting as a low-pass filter, low light at night, lens flare, altered object appearance, snow accumulation, and wet or reflective surfaces (Kerim et al., 2022). These degradations substantially reduce performance for models trained only under standard conditions.

The original empirical study already showed that performance on ACDC drops markedly relative to clear-weather benchmarks and that nighttime is generally the hardest domain (Sakaridis et al., 2021). The same study found that many unsupervised domain adaptation methods effective in synthetic-to-real transfer brought little or no gain over source-only baselines on ACDC, underscoring that real adverse weather is a harder setting than stylized or synthetic shifts (Sakaridis et al., 2021). It also reported that models trained on all conditions, described as “uber” models, outperform ensembles specialized to single conditions, and that nighttime-trained models generalize better to other conditions than vice versa (Sakaridis et al., 2021).

Subsequent work reinforces these conclusions across tasks. In object detection, ACDC serves as a real-world test set demonstrating a pronounced drop in detection performance under adverse weather when detectors are trained only on clear-weather data; data augmentation with Instruct Pix2Pix improved Faster R-CNN robustness in fog, night, rain, and snow, with especially visible gains at night and for pedestrian detection (Gurbindo et al., 13 May 2025). In one reported night setting, AP50 for the walker class increased from 0.415 to 0.500 for Faster R-CNN when augmented data were used (Gurbindo et al., 13 May 2025).

ACDC has also become a reference point in discussions of what dataset correspondence means across modalities. A 2026 LiDAR adaptation paper contrasts ACDC with a new LiDAR dataset suite, noting that ACDC offers real-world adverse-weather data but not frame-to-frame paired clear/adverse LiDAR scans with exact physical correspondence (Anand et al., 11 Apr 2026). This comparison does not diminish ACDC’s role; rather, it clarifies that ACDC’s strength lies in RGB semantic perception with image-level scene correspondence, not in sensor-level one-to-one physical re-renderings.

The dataset’s long-term impact is methodological. Refign uses ACDC correspondences for adaptive pseudo-label correction (Bruggemann et al., 2022). CMA uses them for cross-domain contrastive learning when source data are inaccessible (Bruggemann et al., 2023). FREST uses them to restore adverse features toward normal-condition embeddings (Lee et al., 2024). WeatherProof uses ACDC to test language-guided side information for segmentation under overlapping weather effects (Gella et al., 2024). HeuSCM uses ACDC as one of three widely used benchmarks to validate a learned curriculum for class scheduling in domain-adaptive segmentation (Wang et al., 25 Mar 2026). Across these works, ACDC functions simultaneously as a dataset, a benchmark, and a formalization of adverse-condition perception in which uncertainty, correspondence, and domain shift are all first-class objects.

A plausible implication is that ACDC’s enduring importance stems from the combination of three properties rarely found together: real adverse-weather imagery at scale, dense semantic annotation with explicit invalid-region modeling, and cross-condition image-level correspondences. That combination has made it a persistent testbed not only for measuring robustness, but for defining what robustness should mean in semantic driving scene perception under genuinely adverse visual conditions.