OVRSIS95K: Remote Sensing Mask Dataset

• OVRSIS95K is a balanced remote sensing image-mask dataset that provides 95K image–mask pairs across 35 semantic categories for open-vocabulary segmentation.
• It standardizes annotations via a two-stage semi-automated pipeline with human audits, ensuring high semantic consistency and pixel-accurate masks.
• The dataset underpins OVRSISBenchV2, enabling cross-dataset transfer and robust evaluation in applications like building, road, and flood extraction.

OVRSIS95K is a balanced remote sensing image--mask dataset of about 95K pairs covering 35 common semantic categories across diverse scenes, introduced as the training foundation of OVRSISBenchV2 for open-vocabulary remote sensing image segmentation (OVRSIS) (Li et al., 17 Apr 2026). It was constructed to address fragmented datasets, limited training diversity, and the absence of evaluation protocols that reflect realistic geospatial deployment demands. Within the benchmark, training is performed solely on OVRSIS95K, while evaluation extends to heterogeneous downstream datasets and application protocols such as building extraction, road extraction, and flood detection, thereby positioning OVRSIS95K as a foundational resource for cross-dataset transfer, unseen-class generalization, and application-oriented benchmarking in remote sensing (Li et al., 17 Apr 2026).

1. Definition and problem setting

OVRSIS95K is embedded in the broader problem of open-vocabulary remote sensing image segmentation, where models are trained to segment semantic classes specified by text rather than being restricted to a fixed closed-set label space. In the formulation associated with OVRSISBenchV2, the dataset serves as the sole training source for evaluating cross-dataset generalization over 170K images and 128 categories aggregated from OVRSIS95K and 10 downstream datasets (Li et al., 17 Apr 2026).

The dataset was introduced because OVRSISBenchV1, although useful for exposing the domain gap between natural-image open-vocabulary segmentation models and remote sensing imagery, relied on a single-source training dataset with limited scale and scene diversity. OVRSIS95K directly targets three issues identified in that setting: fragmented datasets, limited training diversity, and realistic evaluation needs. It unifies categories under a single taxonomy, spans five representative remote sensing scene domains, and was designed from the outset to support a multi-domain benchmark with downstream application protocols (Li et al., 17 Apr 2026).

A common source of terminological confusion arises from adjacent literature. OVRSIS95K is unrelated to the OVIS task of open-vocabulary visual instance search, which returns ranked localized image regions from general-image databases in response to textual queries (Liu et al., 2021). It is also distinct from LandDiscover50K, which was introduced for Open-Vocabulary Remote Sensing Image Semantic Segmentation (OVRSISS) and contains 51,846 images covering 40 semantic classes rather than about 95K image--mask pairs covering 35 categories (Ye et al., 2024).

2. Dataset composition and taxonomy

OVRSIS95K contains approximately 95,620 image--mask pairs; in OVRSISBenchV2, the training set count is reported as 94,620 and is rounded to “95K” throughout the paper (Li et al., 17 Apr 2026). The dataset is explicitly described as balanced, with category sampling across five scene domains to alleviate class imbalance and reduce long-tail skew relative to typical remote sensing datasets.

The taxonomy comprises 35 semantic classes:

• airplane
• airport
• baseball field
• basketball court
• bridge
• chimney
• expressway service area
• expressway toll station
• dam
• golf field
• ground track field
• harbor
• overpass
• ship
• stadium
• storage tank
• tennis court
• train station
• vehicle
• windmill
• soccer field
• roundabout
• container crane
• helipad
• building
• road
• water
• tree
• grass
• bareland
• rangeland
• developed space
• agriculture land
• intersection
• background

The five scene domains are town, industrial, forest, waterfront, and wasteland. These domains are characterized, respectively, by dense low-rise housing and mixed-use roads or parking; large rectangular warehouses, storage tanks, and aprons; continuous canopy textures with gaps revealing roads or grass; land--water interfaces with docks, vessels, and breakwaters; and sparse or low vegetation with bare soil or sand, occasionally including sports grounds or scrub (Li et al., 17 Apr 2026).

Masks are delivered at semantic level rather than as the primary instance-level deliverable. The annotation granularity emphasizes pixel-accurate semantics across diverse scales, including small targets such as vehicles and large regions such as developed space. The paper states that images are remote sensing patches drawn from public remote sensing imagery, but does not enumerate specific platforms or sensors. Resolution in the broader benchmark spans 256--4000 pixels on the longer side, and OVRSIS95K is described as focusing on satellite-style top-down content across the five domains (Li et al., 17 Apr 2026).

This composition suggests that OVRSIS95K is intended less as a narrowly curated single-domain corpus than as a standardized semantic substrate for transfer across scene archetypes, resolutions, and downstream task definitions.

3. Construction, annotation, and quality control

OVRSIS95K was built with a two-stage semi-automated pipeline followed by human audit (Li et al., 17 Apr 2026). The first stage is caption-driven category generation. A captioner produces a detailed description for each image, a parser extracts candidate nouns, and these candidates are matched and filtered against the 35-class taxonomy. The second stage is mask generation. A segmentor proposes instance-level masks for the selected categories, after which masks are combined and corrected into semantic-level annotations through merging, boundary correction, and instance filtering.

Human verification constitutes a central component of the dataset construction protocol. The paper reports that 20,000 randomly sampled images across the five domains were checked. For category audit, acceptance was 97.25%, 2.22% were corrected, and 0.53% were discarded as false positives or invalid annotations. For mask audit, acceptance was 91.66%, while 8.34% required correction, specifically including boundary correction, merges, and instance filtering. The false-positive ratio before correction was approximately 0.53%, and all audited annotations were reviewed against a unified taxonomy and cleaned (Li et al., 17 Apr 2026).

The result is described as a scalable pipeline with human verification that yields high semantic consistency and spatial quality while controlling the noise typical of large-scale remote sensing annotation. A plausible implication is that the design prioritizes annotation throughput without relinquishing taxonomy consistency, which is especially important for open-vocabulary evaluation where label ambiguity can otherwise dominate measured generalization.

4. Role within OVRSISBenchV2

OVRSISBenchV2 is built on OVRSIS95K, and the dataset functions as its training foundation (Li et al., 17 Apr 2026). The benchmark aggregates 10 heterogeneous downstream datasets and expands evaluation to 170K images and 128 categories. Relative to OVRSISBenchV1, which unified a smaller collection of datasets and retained training from a single source such as DLRSD or iSAID, OVRSISBenchV2 shifts to a more realistic protocol: train only on OVRSIS95K and evaluate across a much broader multi-domain testbed.

The benchmark includes standard open-vocabulary segmentation and three downstream application protocols:

• building extraction
• road extraction
• flood detection

For building extraction, the evaluation datasets are WHU Aerial, WHUSat-II, Inria, and xBDpre; the model queries the “building” category under a unified taxonomy. For road extraction, the datasets are CHN6-CUG, DeepGlobe roads, Massachusetts roads, and SpaceNet, with the query “road.” For flood detection, the dataset is WBS-SI, with the model querying “water” or using flood-relevant water/background mapping depending on the labeling scheme (Li et al., 17 Apr 2026).

Evaluation is based on per-dataset mean Intersection over Union (mIoU) and mean class accuracy (mACC), with overall averages reported as m-mIoU and m-mACC. Seen and unseen splits are reported where relevant. The benchmark also performs coverage analysis quantifying overlap between OVRSIS95K and each downstream dataset so that unseen semantics remain substantial while some shared classes are retained for stable transfer (Li et al., 17 Apr 2026).

In this structure, OVRSIS95K is not merely a pretraining corpus. It defines the source taxonomy, shapes the seen/unseen interface, and anchors the benchmark’s claim to realistic open-world transfer.

5. Training usage and baseline methodology

The principal baseline introduced alongside OVRSIS95K is Pi-Seg, a model for open-vocabulary remote sensing image segmentation that improves transferability through a positive-incentive noise mechanism (Li et al., 17 Apr 2026). Pi-Seg is implemented atop the CAT-Seg framework and uses CLIP backbones, specifically ViT-B/16 or ViT-L/14, for image and text encoders in dense similarity computation. It incorporates lightweight perturbation modules, cost aggregation blocks, and an upsampling decoder.

The positive-incentive noise mechanism introduces semantically guided, learnable perturbations into both text prototypes and dense visual features. The dense cost map is defined with perturbed embeddings by cosine similarity,

$$C(i,j)=\frac{\langle \hat{U}(i),\hat{I}(j)\rangle}{\|\hat{U}(i)\|_2\|\hat{I}(j)\|_2},$$

where $i$ indexes spatial locations and $j$ indexes class prompts. Segmentation is trained with a per-pixel cross-entropy-style objective over classes. Text perturbation is parameterized by learnable $\mu_{\text{txt}}$ and $\sigma_{\text{txt}}$, while image perturbation is produced by a text-guided cross-attention block that outputs spatially varying $\mu_{\text{vis}}$ and $\sigma_{\text{vis}}$ (Li et al., 17 Apr 2026).

The benchmark uses the unified prompt template “a photo of {class}”. Fixed input sizes are reported as $384\times384$ for ViT-B/16 and $336\times336$ for ViT-L/14. Training uses AdamW, a base learning rate of $2\times10^{-4}$, a cosine schedule, 40K iterations, and batch size 8 on $i$0 NVIDIA H100 GPUs (Li et al., 17 Apr 2026). Pi-Seg avoids the sliding-window inference overhead used by RSKT-Seg for very large inputs, while standard remote sensing tiling and scale-jittering can still be applied for upstream usage of OVRSIS95K.

The paper emphasizes that the primary robustness bottleneck lies in embedding-space fragility rather than in extensive input-level augmentation. This suggests that OVRSIS95K was designed to support methods that rely on stable pixel--text alignment under varied remote sensing conditions, rather than only methods that depend on handcrafted augmentation heuristics.

6. Empirical performance and benchmark effects

On OVRSISBenchV2 with OVRSIS95K training, Pi-Seg ViT-B reports m-mIoU 39.48 and m-mACC 56.90, improving over RSKT-Seg ViT-B at 35.62 and 53.09 and CAT-Seg ViT-B at 37.88 and 55.72. Pi-Seg ViT-L reports m-mIoU 44.40 and m-mACC 63.16, improving over RSKT-Seg ViT-L at 40.10 and 59.94, CAT-Seg ViT-L at 42.49 and 61.01, and OVRS ViT-L at 42.77 and 59.31 (Li et al., 17 Apr 2026). The paper highlights notable per-dataset values including UAVid mIoU 59.61 and Vaihingen mIoU 31.70 for Pi-Seg ViT-B.

On downstream tasks with OVRSIS95K training, Pi-Seg ViT-B achieves building extraction mIoUs including WHU Aerial 85.88 and Inria 78.61, and Pi-Seg ViT-L further improves in most cases. For road extraction, Pi-Seg ViT-B yields the best or near-best means, and for flood detection on WBS-SI the model shows clear gains in delineating water and flooded regions under open-vocabulary prompts (Li et al., 17 Apr 2026).

The paper also reports OVRSISBenchV1 comparisons to show transfer under seen and unseen splits. For UAVid, Pi-Seg achieves seen mIoU 66.37 and unseen mIoU 24.30, with seen mACC 78.70 and unseen mACC 49.83. These values are presented as outperforming alternatives in unseen performance (Li et al., 17 Apr 2026).

Two consequences follow directly from these results. First, benchmark difficulty rises materially when OVRSIS95K is paired with a broader evaluation suite, implying that the dataset’s value lies partly in the realism of the transfer protocol it enables. Second, balanced training diversity appears to matter at least as much as architectural choice, since the benchmark is explicitly framed as a response to the limited scope of earlier single-source training setups.

7. Relation to adjacent datasets, misconceptions, and limitations

OVRSIS95K occupies a specific position within remote sensing open-vocabulary segmentation research. It is larger than LandDiscover50K, which provides 51,846 images and 40 semantic classes for OVRSISS and is associated with the GSNet architecture rather than Pi-Seg (Ye et al., 2024). LandDiscover50K emphasizes harmonization across datasets such as OEM, LoveDA, DeepGlobe Land Cover, SIOR, and SOTA, whereas OVRSIS95K is described as a balanced dataset spanning five representative scene domains and explicitly built to underpin OVRSISBenchV2 (Li et al., 17 Apr 2026).

A separate misconception concerns confusion with the visually similar acronym OVIS. OVIS, introduced for general-image open-vocabulary visual instance search, is a region retrieval problem in which a textual query returns ranked visual instances rather than semantic masks (Liu et al., 2021). OVRSIS95K, by contrast, belongs to remote sensing semantic segmentation and supplies image--mask pairs rather than a region-ranking database.

The limitations reported for OVRSIS95K and its surrounding benchmark are also specific. Category semantics are primarily short prompts, so fine-grained distinctions such as ship versus vehicle versus airplane can remain ambiguous. Sensor or platform metadata and geography are not enumerated, making per-sensor domain-shift analysis indirect. Masks are semantic-level rather than panoptic instance annotations (Li et al., 17 Apr 2026). These constraints do not negate the dataset’s utility, but they delimit the kinds of generalization claims that can be made directly from benchmark outcomes.

Access points are provided through public dataset and code repositories, including Hugging Face entries for OVRSIS95K and OVRSISBench test aggregations and a GitHub repository for the Pi-Seg implementation. Licensing terms and usage restrictions are not specified in the paper; consultation of the dataset cards and repository is identified as the authoritative route for license information (Li et al., 17 Apr 2026).