--- title: WeatherProof Dataset Challenge url: https://www.emergentmind.com/topics/weatherproof-dataset-challenge type: topic --- # WeatherProof Dataset Challenge to=arxiv_search.search 天天中彩票微信 彩神争霸高 "query": "\"WeatherProof\" semantic segmentation adverse weather challenge", "max_results": 10} to=arxiv_search.search 天天中彩票和 񹚊ppion 一级a做爰片{ "query": "\"WeatherProof Dataset Challenge\"", "max_results": 10 } to=arxiv_search.search 天天中彩票无法 "query": "\"A Paired-Dataset Approach to Semantic Segmentation in Adverse Weather\" OR \"Leveraging Language Guidance for Semantic Segmentation in Adverse Weather\" OR \"Technical Report for CVPR 2024 WeatherProof Dataset Challenge\"", "max_results": 10 } The WeatherProof Dataset Challenge denotes a set of adverse-weather vision benchmarks whose central formulation is semantic segmentation on paired clear and weather-degraded images of the same scene. In its best documented form, WeatherProof is a paired semantic segmentation dataset introduced to measure and reduce the robustness gap between clear and adverse-weather perception, covering rain, fog, and snow with high-quality pixel-level labels over 10 semantic classes [2312.09534]. The name is most closely associated with the CVPR 2024 UG2+ Track 3 challenge and later semantic-segmentation reports built on the same paired-data paradigm, although some later literature also applies “WeatherProof” colloquially to a distinct 2026 video weather-removal challenge [2407.01579][2605.22216][2604.10655]. ## 1. Historical formation and scope WeatherProof emerged from the observation that strong semantic segmentation models trained on clear-weather benchmarks exhibit a large performance drop on images degraded by rain, fog, or snow. The dataset paper presents WeatherProof as “the first semantic segmentation dataset with accurate clear and adverse weather image pairs,” explicitly designed so that clear and degraded frames share the same underlying scene and the same semantic labels [2312.09534]. A closely related paper then framed the benchmark around language-guided segmentation, again emphasizing that paired clear and adverse images make it possible to attribute performance differences to weather rather than scene change [2403.14874]. The CVPR 2024 UG2 WeatherProof Dataset Challenge operationalized this setting as a competition in robust semantic segmentation on paired real images captured under varying adverse weather conditions. The 2024 technical reports describe the dataset as containing scenes “from all around the world,” degraded by varying degrees of weather, with paired “clean vs. degraded” imagery used to support controlled training and evaluation [2407.01579]. A later CVPR 2026 8th UG2+ Challenge Track 2 report describes the same benchmark line as a paired semantic segmentation dataset in which clean images are labeled and degraded-weather images can be exploited as unlabeled data for semi-supervised training [2605.22216]. The label “WeatherProof Dataset Challenge” is not completely uniform across the literature. A 2021 autonomous-driving paper uses the term for a weather and light-level classification benchmark in RCCC imagery rather than semantic segmentation [2104.14042]. A 2026 LoViF report states that “WeatherProof” refers only colloquially to the “LoViF 2026 Challenge on Weather Removal in Videos,” whose official dataset name is WRV [2604.10655]. In current adverse-weather segmentation research, however, WeatherProof is primarily the paired semantic-segmentation benchmark introduced in 2023 and used in the UG2+ challenge reports [2312.09534]. ## 2. Dataset design, pairing, and annotation The defining property of WeatherProof is pairwise alignment: each adverse-weather image is paired with a clear-weather image of the same scene with “very minimal differences,” so that both frames can share identical semantic labels [2312.09534]. This pairing is the basis for both controlled evaluation and paired training. The dataset uses real adverse-weather imagery rather than synthetic weather, drawing from GT-RAIN and WeatherStream, and covers rain, snow, and fog or haze across diverse geographic locations, camera parameters, and resolutions [2312.09534][2403.14874]. The paired dataset papers describe a 10-class taxonomy: - background - tree - structure - road - terrain-snow - terrain-grass - terrain-other - stone - building - sky These classes are annotated with high-quality pixel-level masks, with emphasis on accurate object borders and minimizing spurious background gaps between adjacent objects [2312.09534]. The large-scale dataset description reports over 174.0K images in total, with 147.8K training images and 26.2K testing images [2312.09534]. A later challenge report presents the same dataset as “>174K” images, partitioned into 147.8K training and 26.2K test images [2605.22216]. Challenge reports also describe the corpus by scene counts. The 2024 winning solution refers to 513 scenes for training and 38 scenes for validation, and distinguishes between WeatherProof, WeatherProofClean, and an auxiliary expansion called WeatherProofExtra [2406.03799]. WeatherProofClean consists of the clean images paired with the degraded scenes and shares the same annotations, while WeatherProofExtra contains 160 scenes with a wider viewing angle and is used to address the reported gap between training and test imagery in resolution and field of view [2406.03799]. A narrower class list appears in one 2024 submission report because its results table gives per-class IoUs only for nine foreground categories: building, structure, road, sky, stone, terrain-grass, terrain-other, terrain-snow, and tree [2407.01579]. This does not alter the 10-class dataset taxonomy reported in the benchmark papers; it reflects the class reporting used in that particular submission. ## 3. Evaluation protocols and challenge structure The official metric for the semantic-segmentation challenge reports is mean Intersection over Union. The metric is defined as $$ \mathrm{IoU}_c = \frac{\mathrm{TP}_c}{\mathrm{TP}_c + \mathrm{FP}_c + \mathrm{FN}_c}, \qquad \mathrm{mIoU} = \frac{1}{C} \sum_{c=1}^{C} \mathrm{IoU}_c. $$ This formulation is stated explicitly in the 2024 second-place report and repeated, with related segmentation metrics, in the 2026 semi-supervised report [2407.01579][2605.22216]. The 2026 report additionally tracks Pixel Accuracy, Mean Accuracy, frequency-weighted IoU, Dice per class, and mean Dice for analysis, but identifies mIoU as the primary challenge quantity [2605.22216]. A distinctive evaluation idea in the WeatherProof benchmark papers is paired testing on both adverse and clear frames of the same scene. This supports direct measurement of a weather-induced robustness gap, $\Delta = \mathrm{mIoU}(\text{clear}) - \mathrm{mIoU}(\text{adverse})$, rather than conflating weather degradation with scene novelty [2312.09534]. That paired protocol is central to the benchmark’s methodological significance: it decouples semantic variation from photometric and atmospheric degradation. The challenge constraints reported in solution papers vary by year and track. One 2024 report states that no additional or external datasets were used, and that the validation set was not folded into training for final testing [2407.01579]. By contrast, the 2024 winning report notes that, during final testing, all available WeatherProof data, including train and validation, were used for training [2406.03799]. The 2026 semi-supervised report again states that it used only WeatherProof data and adhered to a “no external data” policy, employing a 9:1 random split of the official training set into train and validation [2605.22216]. These reports show that WeatherProof challenge practice has combined a stable metric with track-specific decisions about data usage. ## 4. Core methodological paradigms The benchmark literature converges on a small number of technical paradigms. The original WeatherProof training paper proposes paired training on clear and adverse frames from the same scene, with standard segmentation losses applied to both views and two explicit consistency terms: a Feature Consistency Loss that aligns encoder features across the pair, and an Output Consistency Loss that aligns decoder predictions [2312.09534]. The same work introduces CLIP-based language guidance, in which weather text embeddings are combined into a composition vector and injected through cross-attention. This paired-data pipeline is reported to improve adverse-weather mIoU by up to 18.4% relative to adverse-only training [2312.09534]. A related language-guidance paper refines this idea by using CLIP text prompts to represent adverse-weather composition and injecting the resulting side information into segmentation backbones such as InternImage XL, ConvNeXt XL, and Swin Transformer [2403.14874]. That paper emphasizes composite weather effects such as rain+fog and snow+fog, and reports that language guidance substantially reduces the performance penalty between single-effect and multi-effect weather cases [2403.14874]. Challenge solutions extend these benchmark ideas in several directions. The 2024 winning system is an ensemble of five model families, including UPerNet-depth and SETRMLA-depth with Depth Anything-pretrained DINOv2 backbones, InternImage with enhanced CLIP guidance that uses both weather and category prompts, and OneFormer with a large-scale Swin backbone [2406.03799]. It further introduces WeatherProofExtra, adverse-weather watermark augmentation, DAT-based super-resolution, multi-scale inference, and a final per-pixel majority-vote ensemble [2406.03799]. The 2024 second-place report takes a different route: InternImage-H is used as a pre-trained large-scale vision foundation model, Mask2Former provides the segmentation head, and clean, degraded, and denoised images generated by DA-CLIP are mixed during training to expose the model to different noise levels [2407.01579]. The final system combines multiple models through voting, then applies dense CRF and morphological transformations as post-processing [2407.01579]. A fourth-place 2024 solution introduces a two-stage pipeline in which the test images are concatenated into a video sequence and processed with LLRT, a low-rank video deraining method, to produce “high-fidelity pseudo ground truths,” after which InternImage-H is used for segmentation [2406.05513]. The 2026 semantic-segmentation report moves instead toward semi-supervision: UniMatch V2 is adapted with a frozen DINOv2-Base encoder and a DPT head, clean images are treated as labeled data, and all degraded images are treated as unlabeled data for teacher-generated pseudo-label training with confidence thresholding and test-time augmentation [2605.22216]. ## 5. Representative results and leaderboard outcomes The WeatherProof challenge literature reports results at several levels: benchmark ablations, challenge leaderboard placements, and later track extensions. The table below summarizes prominent reported outcomes. | Setting | Reported result | Notes | |---|---:|---| | InternImage adverse-only on WeatherProof | 43.32 mIoU | Baseline in paired-training paper [2312.09534] | | InternImage paired + losses + CLIP | 51.31 mIoU | Adverse-weather test [2312.09534] | | InternImage baseline on WeatherProof | 46.47 mIoU | WeatherProof adverse test [2403.14874] | | InternImage + language guidance | 51.19 mIoU | WeatherProof adverse test [2403.14874] | | CVPR 2024 winning solution | 0.47 mIoU | 1st place [2406.03799] | | CVPR 2024 second-place solution | 45.10 mIoU | 2nd place [2407.01579] | | CVPR 2024 two-stage LLRT + InternImage | 0.43 mIoU | 4th place [2406.05513] | | CVPR 2026 UniMatch V2-style pipeline | 0.80 mIoU | Semi-supervised + TTA [2605.22216] | The 2024 second-place report is unusually explicit about incremental gains. Its baseline InternImage-H + Mask2Former system achieved 39.61 mIoU, the addition of ensemble voting increased performance to 43.60 mIoU, dense CRF further raised it to 44.48 mIoU, and morphology produced a final 45.10 mIoU [2407.01579]. This report also gives per-class IoUs for each stage and notes that the result was obtained with fewer submissions than the other winners [2407.01579]. The 2024 winning report shows a different profile. Single-model results include 0.43 mIoU for InternImage, 0.44 for InternImage + CLIP, 0.42 for UPerNet-depth, 0.46 for SETRMLA-depth, and 0.43 for OneFormer, while the six-output ensemble reaches 0.47 mIoU and ranks first on the final leaderboard [2406.03799]. Its prompt ablation further reports mIoU values of 0.534 for category prompts only, 0.535 for weather prompts only, and 0.552 for combined category and weather prompts on a WeatherProof train/evaluate split [2406.03799]. The benchmark papers contextualize these challenge outcomes. On WeatherProof adverse-weather test images, the paired-data paper reports that InternImage rises from 43.32 mIoU under adverse-only fine-tuning to 45.24 with paired training and 51.31 with paired training, consistency losses, and CLIP guidance [2312.09534]. The language-guidance paper reports a comparable pattern, with InternImage moving from 46.47 to 51.19 mIoU on WeatherProof adverse frames and from 54.44 to 57.14 on paired clear frames [2403.14874]. On ACDC, the same approach raises InternImage from 76.17 to 82.6 mIoU [2403.14874]. The 2026 semantic-segmentation report indicates a sharp shift in the reported operating point for that later track: clean-only supervised training gives 0.69 mIoU and 0.69 mDice, adding degraded images through semi-supervised training raises both to 0.79, and test-time augmentation lifts both to 0.80 [2605.22216]. Since this is a later UG2+ track, these values should be read as belonging to a different challenge instantiation rather than as directly comparable to the 2024 leaderboard numbers. ## 6. Limitations, controversies, and related extensions Several limitations recur across the WeatherProof literature. The paired-data formulation assumes minimal differences between clear and adverse frames; the benchmark papers note that severe occlusions and substantial dynamics can break the assumptions behind consistency losses and label sharing [2312.09534][2403.14874]. Challenge reports also document incomplete reproducibility. The 2024 second-place report does not disclose dense-CRF parameter settings, exact architectural tweaks, or several inference details, and explicitly notes that the lack of such hyperparameters may hinder precise reproduction [2407.01579]. The two-stage LLRT report omits solver details, alignment procedures, and ablations quantifying the effect of deraining on convergence [2406.05513]. Another recurrent theme is that larger models are not always better. The second-place 2024 report states that InternImage-XL performed significantly worse than InternImage-H in that challenge setting [2407.01579]. The same report also observes strong category-wise fluctuations across model versions and a pronounced dependence on post-processing; dense CRF and morphology materially improve performance, suggesting that the base segmentation output can remain fragmented or uncertain under heavy degradation [2407.01579]. The term “WeatherProof Dataset Challenge” itself is a source of possible confusion. In the semantic-segmentation literature it refers to the paired clear/adverse-weather benchmark and its associated UG2+ challenge tracks [2312.09534][2407.01579]. In a different adverse-weather line, the LoViF 2026 “Challenge on Weather Removal in Videos” is described as “WeatherProof” only colloquially; its official dataset is WRV, its task is single-track video weather removal, and its leaderboard score is $\mathrm{Final\_Score} = \mathrm{PSNR}_Y + 10\,\mathrm{SSIM}_Y - 5\,\mathrm{LPIPS} - 30\,\mathrm{Warp\_Error}$ [2604.10655]. A still earlier autonomous-driving paper uses the label for a weather, light-level, and street-type classification dataset built from RCCC camera data rather than for segmentation [2104.14042]. Taken together, the literature presents WeatherProof less as a single frozen competition than as a benchmark lineage organized around a specific technical idea: controlled adverse-weather learning through paired or aligned data. Its major contributions are the separation of weather effects from scene variation, the promotion of robustness-oriented evaluation via mIoU and paired clear/adverse testing, and the demonstration that paired training, language guidance, semi-supervision, and carefully engineered ensembling can all materially improve segmentation under rain, fog, and snow [2312.09534][2403.14874][2605.22216]. Source: https://www.emergentmind.com/topics/weatherproof-dataset-challenge