Region-Based Active Learning

• Region-based active learning is an approach where queries target localized regions rather than entire samples, leveraging spatial structure for efficient annotation.
• It reduces redundancy by selecting informative subregions, thus minimizing annotation effort in dense prediction tasks across various domains.
• Empirical results demonstrate that region-based methods nearly match full-supervision performance while significantly lowering annotation cost.

Searching arXiv for recent and foundational papers on region-based active learning. Region-based active learning is a family of active learning methods in which the query unit is a region rather than an entire image, scan, graph node considered in isolation, or a single point in a pool. In dense prediction, the region is typically a fixed-size crop, a rectangular subregion, a square box, a sub-scene region, or a region of interest; in graph and metric-space formulations, it may be a neighborhood, a ball, an interval, a box, or a topological region. Across these settings, the recurring motivation is that informative content is localized, annotation cost is spatially nonuniform, and nearby candidates are often redundant, so a label budget is used more effectively when active learning selects where inside the input space annotation should occur (Mackowiak et al., 2018, Colling et al., 2020, Dasgupta et al., 2023, Abel et al., 2019).

1. Conceptual scope and historical development

Region-based active learning emerged as a response to a mismatch between dense annotation cost and sample-level query granularity. In semantic segmentation, the early formulation in “CEREALS - Cost-Effective REgion-based Active Learning for Semantic Segmentation” treats active learning as a pool-based, batch-mode process in which fixed-size square windows are selected instead of full images, explicitly combining information measures with annotation-cost estimation and non-overlap constraints (Mackowiak et al., 2018). In parallel, “Adaptive Region-Based Active Learning” defines region-based active learning in a more abstract way: the learner adaptively partitions the input space into disjoint regions $X_1,\dots,X_K$, learns a separate predictor per region, and returns a piecewise predictor

$h(x)=\sum_{k=1}^K 1_{x\in X_k} h_k(x),$

thereby competing with a richer region-wise comparator class than a single global predictor (Cortes et al., 2020).

Subsequent work broadened the idea in two directions. One direction specialized the region unit to dense vision tasks, where region-based active learning became a method for selecting crops, patches, or sub-scene regions for semantic segmentation, medical segmentation, point cloud segmentation, and object detection (Colling et al., 2020, Wu et al., 2021, Liou et al., 2023, Laielli et al., 2021). The other direction generalized the notion of region itself. “Active learning using region-based sampling” defines regions as neighborhoods or balls in a metric space and identifies overlap areas of oppositely biased regions as “known unknowns” that should attract future queries (Dasgupta et al., 2023). “Active Learning with Simple Questions” formalizes region queries as pairs $(T,z)$, where the learner asks whether every example in a region $T$ has label $z$, and measures query-language complexity via the VC dimension of the family of regions (Kontonis et al., 2024). “Pool-Based Active Learning with Proper Topological Regions” further defines regions through density-adaptive persistent clustering, so that standard pointwise acquisition rules can be lifted to topological regions with label propagation inside each region (Hadjadj et al., 2023).

A consistent historical pattern is that region-based active learning has been developed when sample-level active learning was viewed as too coarse and pointwise active learning as too fine-grained or too sparse. Several papers make that contrast explicit: image-level active learning spends budget on redundant or easy content, while pixel-level or point-level active learning may be impractical, lose local context, or create incomplete-label effects; region-based querying is presented as the middle ground (Cai et al., 23 Jul 2025, Laradji et al., 2020, Laielli et al., 2021).

2. Formal formulations of regions and queries

In dense semantic segmentation, a standard formulation begins with a pool $\mathcal P$, a labeled subset $\mathcal L$, and an unlabeled subset $\mathcal U$. MetaBox+ models the segmentation network as

$$f : [0,1]^{w \times h \times 3} \to [0,1]^{w \times h \times c},$$

defines a pixel-wise priority map

$$g: [0,1]^{w \times h \times c} \to [0,1]^{w \times h},$$

and aggregates it over a set $h(x)=\sum_{k=1}^K 1_{x\in X_k} h_k(x),$0 of fixed-size square boxes $h(x)=\sum_{k=1}^K 1_{x\in X_k} h_k(x),$1 by summation, producing region scores that are then ranked under a non-overlap constraint (Colling et al., 2020). CEREALS uses the same basic pixel-to-region logic: uncertainty or disagreement is first computed per pixel, then aggregated over sliding square windows, normalized, fused with a cost map, and filtered by non-maximum suppression (Mackowiak et al., 2018). RIPU instead defines a candidate region around every pixel center as

$h(x)=\sum_{k=1}^K 1_{x\in X_k} h_k(x),$2

and scores that neighborhood by multiplying region impurity and prediction uncertainty (Xie et al., 2021).

In adaptive input-space partitioning, the region is not a fixed crop but an evolving subset of the domain. ARBAL starts from a single region $h(x)=\sum_{k=1}^K 1_{x\in X_k} h_k(x),$3, considers axis-aligned binary splits

$h(x)=\sum_{k=1}^K 1_{x\in X_k} h_k(x),$4

and performs a split only when the empirical reduction in region risk exceeds a threshold $h(x)=\sum_{k=1}^K 1_{x\in X_k} h_k(x),$5 with a concentration slack. Region formation is thus an explicit model-selection step driven by queried labels (Cortes et al., 2020).

In metric-space formulations, the region is a neighborhood family $h(x)=\sum_{k=1}^K 1_{x\in X_k} h_k(x),$6 over the unlabeled pool $h(x)=\sum_{k=1}^K 1_{x\in X_k} h_k(x),$7. Region-based sampling groups balls by level according to their sample-set size, estimates whether each ball has significantly positive, significantly negative, or insignificant average label bias, and assigns a point the provisional label $h(x)=\sum_{k=1}^K 1_{x\in X_k} h_k(x),$8, $h(x)=\sum_{k=1}^K 1_{x\in X_k} h_k(x),$9, $(T,z)$0, or “!” depending on which biased regions contain it. The symbol “!” marks exactly the case in which overlapping neighborhoods support opposite labels (Dasgupta et al., 2023). In the query-LLM of region queries, a region query is

$(T,z)$1

and the answer is whether every example in $(T,z)$2 has label $(T,z)$3, where $(T,z)$4 is the labeler’s domain; this shifts the unit of interaction from labeling a point to certifying homogeneity over a region (Kontonis et al., 2024).

These formulations share a common structural move: they replace scalar sample scoring with region-aware scoring or region-aware querying. The specific mechanisms differ—aggregation over boxes, adaptive partitions, neighborhood bias estimation, or set-valued homogeneity queries—but the operational question becomes which region should be labeled or certified next, not simply which sample.

3. Region definitions across application domains

Region-based active learning is not tied to a single geometric primitive. The choice of region is strongly task-dependent.

Setting	Region definition	Representative work
2-D semantic segmentation	Fixed square boxes, grid regions, or local neighborhoods	(Mackowiak et al., 2018, Colling et al., 2020, Xie et al., 2021)
Medical image segmentation	Non-overlapping rectangular grid cells or selected image regions	(Laradji et al., 2020, Qiu et al., 8 Dec 2025)
Point cloud segmentation	VCCS sub-scene regions	(Wu et al., 2021)
Object detection	Multi-scale crops or arbitrary rectangular regions	(Liou et al., 2023, Laielli et al., 2021)
Graph and metric-space learning	First-hop neighborhoods, balls, intervals, boxes, PTR	(Abel et al., 2019, Dasgupta et al., 2023, Hadjadj et al., 2023, Kontonis et al., 2024)

In semantic segmentation, fixed square regions are dominant because they are easy to enumerate, score, and annotate. CEREALS studies $(T,z)$5, $(T,z)$6, and $(T,z)$7 windows and reports that smaller windows work better (Mackowiak et al., 2018). MetaBox+ uses fixed side length $(T,z)$8 and queries $(T,z)$9 non-overlapping boxes per active-learning round on Cityscapes (Colling et al., 2020). The spatial-diversity framework divides each image into non-overlapping fixed-size regions, with $T$0 for Cityscapes and $T$1 for PASCAL VOC 2012 (Cai et al., 23 Jul 2025). RIPU uses overlapping square neighborhoods centered at each pixel, with $T$2 for region-based annotating and $T$3 for its pixel-based variant (Xie et al., 2021).

In medical segmentation, regions are often matched to annotation workflow. The COVID-19 CT method divides each slice into $T$4 equally sized non-overlapping rectangles and ranks them by mean per-pixel entropy, while requesting point-level supervision inside the selected region instead of dense masks (Laradji et al., 2020). In DECOMP, the region remains a fixed-size square patch for segmentation and a predefined ROI for ROI classification, but selection is conditioned on pseudo-label decomposition into class-specific components (Qiu et al., 8 Dec 2025).

In 3-D point cloud semantic segmentation, ReDAL abandons fixed image crops and over-segments each scan with VCCS to produce medium-sized annotation-friendly sub-scenes. The method is explicit that these regions are not intended to be semantically pure segments; they are intended to preserve local geometric structure while remaining practical annotation units (Wu et al., 2021).

In object detection, region design mediates between whole-image and instance-level labeling. MuRAL defines a region as a crop $T$5, generated by sliding a fixed-size window over multi-scale versions of the image so that the effective perception field changes with scale (Liou et al., 2023). ReAL constructs a region around a predicted object, adds nearby but feature-dissimilar neighbors, and then queries the minimum bounding rectangle containing all selected predicted boxes (Laielli et al., 2021).

In graph and general metric spaces, a “region” is purely relational. Regional graph active learning uses the first-neighbor set $T$6 as the region around node $T$7 (Abel et al., 2019). Region-based sampling over metric spaces uses a family of balls or intervals at multiple scales (Dasgupta et al., 2023). Proper Topological Regions are clusters produced by ToMATo on a density-adaptive $T$8-Rips graph, with label propagation from the highest-density point in each region (Hadjadj et al., 2023).

This variety suggests that “region-based” denotes a query granularity rather than a particular geometry. The region may be fixed, adaptive, learned, topological, or graph-defined; what is constant is that active learning exploits local structure rather than treating the input space as an undifferentiated set of whole samples.

4. Acquisition criteria, diversity, and cost-awareness

The earliest region-based methods in dense prediction largely inherited uncertainty sampling, but the literature quickly moved toward richer acquisition criteria. CEREALS uses pixel-wise entropy or vote entropy, aggregates them over candidate windows, predicts a dense click-cost map with a separate cost network, and combines information and cost through fusion functions

$T$9

with the multiplicative $z$0 giving the best practical result with estimated costs (Mackowiak et al., 2018).

MetaBox+ replaces pixel-wise uncertainty with predicted segment quality. A meta regressor estimates the segment-wise $z$1 of each predicted segment from hand-crafted features, broadcasts that predicted quality to all pixels in the segment, inverts it by

$z$2

and combines it with an RDP-based click-cost priority map

$z$3

through

$z$4

The motivation is that entropy concentrates on contours, while poor predictions in semantic segmentation often concern entire connected components (Colling et al., 2020).

RIPU introduces a different kind of region statistic. For each neighborhood $z$5, it computes the entropy of pseudo-label class proportions,

$z$6

and multiplies it by region uncertainty

$z$7

yielding the acquisition score

$z$8

Here the region is informative not merely because predictions are uncertain, but because a locally impure pseudo-label composition indicates boundaries or multi-class neighborhoods (Xie et al., 2021).

A separate line of work emphasizes redundancy control. “Exploring Spatial Diversity for Region-based Active Learning” argues that when the candidate pool consists of local regions, nearby selections are often redundant, and therefore diversity should be modeled spatially, not only in feature space. Its Max-Min objective combines uncertainty with a piece-wise spatial distance that penalizes nearby regions in the same image while not over-rewarding arbitrarily distant ones (Cai et al., 23 Jul 2025). ReDAL uses a different diversity mechanism: region informativeness is

$z$9

combining entropy, color discontinuity, and structural complexity, then a cluster-wise decay rule

$\mathcal P$0

downweights repeated selection from similar region clusters (Wu et al., 2021).

DECOMP shifts diversity from feature geometry to predicted class decomposition. It estimates class-wise confidence

$\mathcal P$1

defines weights

$\mathcal P$2

scores images by weighted class presence,

$\mathcal P$3

and then samples regions class-conditionally from pseudo-label components (Qiu et al., 8 Dec 2025). MuRAL applies a closely related principle to object detection, where the region score is the mean of $\mathcal P$4confidence over predicted objects in a crop, reweighted by inverse class frequency,

$\mathcal P$5

followed by a round-robin scale-aware selection policy (Liou et al., 2023).

A different strategy is to learn the acquisition rule itself. Reinforced active learning for image segmentation represents each candidate region with pooled entropy, predicted class distribution, and KL-divergence histograms to labeled and unlabeled sets, then trains a modified DQN to select batches of regions whose annotation leads to the largest downstream gain in mean IoU per class (Casanova et al., 2020). This suggests that region-based active learning can be framed either as explicit scoring or as sequential decision-making over region descriptors.

5. Empirical performance across major tasks

In semantic segmentation, region-based active learning repeatedly outperforms whole-image querying under the metrics used in the respective papers. CEREALS reports that the learned cost-aware multiplicative fusion reaches 95% of full-supervision performance on Cityscapes with about 17% of annotation effort in clicks, whereas image-based entropy requires substantially more budget in both pixels and clicks (Mackowiak et al., 2018). MetaBox+ reports that it achieves 95% of the full-data mIoU using $\mathcal P$6 annotation effort for DeepLabv3+ and $\mathcal P$7 for FCN8 under the click-based metric $\mathcal P$8, with results framed around the reduction of annotation effort rather than labeled area alone (Colling et al., 2020). The spatial-diversity framework reports 95% of fully supervised performance with $\mathcal P$9 labeled pixels on Cityscapes and $\mathcal L$0 on PASCAL VOC 2012, and also shows that Entropy+Spatial beats RAL while using far less annotation budget in a fixed-budget Cityscapes comparison (Cai et al., 23 Jul 2025).

In domain-adaptive semantic segmentation, RIPU reports that region-based annotating with only $\mathcal L$1 target queried regions or pixels nearly reaches supervised performance on GTAV $\mathcal L$2 Cityscapes and SYNTHIA $\mathcal L$3 Cityscapes, and clearly exceeds both image-based and point-based active baselines under the budgets used in that work (Xie et al., 2021). Reinforced active learning for image segmentation reports that on Cityscapes its DQN-based region policy requires roughly 30% less additional labeled data than the most competitive baseline to reach the same performance, and that with 20k selected regions—about 6% of total additional pixels from the unlabeled pool—it reaches 64.5% mIoU, about 96% of the fully supervised performance reported in that study (Casanova et al., 2020). DECOMP extends the empirical scope further: it reports that on Cityscapes 95% of the full-annotation benchmark is reached with less than 2.5% of total image area annotated, and on KiTS23 95% is reached with less than 0.15% of dataset volume annotated (Qiu et al., 8 Dec 2025).

In medical segmentation, the COVID-19 CT work shows that labeling small regions is more efficient than labeling whole images and that entropy-based region ranking beats random region ranking. Its headline number is that on COVID-19-A-Mixed, weakly supervised entropy-based region active learning reaches about 90% of full-supervision Dice with only 7% of the labeling effort (Laradji et al., 2020). A different medical direction appears in breast MRI, where the “region-based” element is anatomy-aware whole-sample selection rather than local patch querying: the Nearest Point strategy combined with 30% of the training data is reported as the best balance between segmentation performance, efficiency, and environmental sustainability (Narimani et al., 5 Jan 2026). This suggests a broader usage of “region-based” in which the geometry of the target anatomical region drives acquisition, even if the queried unit remains the whole case.

In 3-D point cloud segmentation, ReDAL reports that it achieves the performance of 90% fully supervised learning while less than 15% and 5% annotations are required on S3DIS and SemanticKITTI, respectively, and that on SemanticKITTI with MinkowskiNet it reaches 95% of the fully supervised result with only 4% labeled points (Wu et al., 2021). The paper’s ablations further state that changing the query unit from scans to regions is the largest improvement, and that diversity-aware selection is essential once the query unit becomes local.

In object detection, region-based methods are motivated by the instability of instance-only selection and the redundancy of image-level selection. MuRAL reports that it surpasses all existing coarse-grained and fine-grained baselines on Cityscapes and MS COCO, with especially strong gains on difficult categories and with the multi-scale, scale-aware design identified as the dominant factor in ablation (Liou et al., 2023). ReAL, targeting cluttered overhead imagery, reports up to 50–60% reduction in labeling effort at matched performance relative to image-level and object-level baselines, and emphasizes improvements in rare-object search on xView (Laielli et al., 2021).

Across these tasks, a recurring empirical conclusion is that region-level acquisition changes not only label efficiency but also class coverage. Several papers explicitly report stronger performance on under-represented classes, small objects, or difficult categories when acquisition is localized and redundancy-aware (Casanova et al., 2020, Xie et al., 2021, Liou et al., 2023, Qiu et al., 8 Dec 2025).

6. Theory, limitations, and open problems

The theoretical literature shows that region-based active learning is not only a heuristic for dense vision. ARBAL proves that if the algorithm splits into $\mathcal L$4 regions, then with high probability the final risk satisfies

$\mathcal L$5

so the benchmark itself improves when a useful split is made (Cortes et al., 2020). Region-based sampling over metric spaces proves finite-pool label complexity guarantees without smoothness assumptions, expressing the complexity of correctly labeling a point through critical scales $\mathcal L$6, $\mathcal L$7, and multiscale boundary sets $\mathcal L$8 (Dasgupta et al., 2023). Region queries with “simple questions” prove that for any hypothesis class of VC dimension $\mathcal L$9, there exists a region query family of VC dimension $\mathcal U$0 such that every $\mathcal U$1-point pool can be perfectly labeled with $\mathcal U$2 region queries, together with a matching lower-bound phenomenon in the worst case (Kontonis et al., 2024). PTR adds a topological perspective in which region construction itself is the key object of study: a density-adaptive $\mathcal U$3-Rips graph, ToMATo clustering, and a region-level propagation rule define Proper Topological Regions that can wrap standard pointwise acquisition strategies (Hadjadj et al., 2023).

At the same time, the empirical literature is explicit about limitations. MetaBox+ requires an additional fully labeled meta-training subset $\mathcal U$4, retrains the meta regressor every active-learning round, uses a heuristic RDP-based cost estimator, and restricts acquisition to fixed square boxes (Colling et al., 2020). RIPU depends on pseudo-label quality under domain shift and on hand-designed square neighborhoods (Xie et al., 2021). DECOMP depends on pseudo-label decomposition and thresholded confidence, and class coverage remains probabilistic rather than guaranteed (Qiu et al., 8 Dec 2025). ReDAL depends on VCCS over-segmentation and measures budget by point count rather than actual human time (Wu et al., 2021). ReAL and MuRAL both rely on detector predictions to define candidate regions, so poor early detectors can degrade region generation (Laielli et al., 2021, Liou et al., 2023).

Several common misconceptions are corrected by this literature. First, region-based active learning is not simply pixel-level active learning with larger units. Multiple papers explicitly distinguish image-level, pixel-level or point-level, and region-based regimes, treating region-based acquisition as a separate design point that preserves context while reducing redundancy (Cai et al., 23 Jul 2025, Laradji et al., 2020, Xie et al., 2021). Second, labeled area is not an adequate proxy for annotation effort in dense labeling. CEREALS and MetaBox+ both argue that click-based or interaction-based cost accounting can differ sharply from pixel fraction because boundaries and polygon complexity dominate human effort (Mackowiak et al., 2018, Colling et al., 2020). Third, diversity in region-based active learning cannot be treated as a purely feature-space concept. Spatial diversity, batch redundancy, and coverage of low-confidence or minority classes recur as independent design concerns (Cai et al., 23 Jul 2025, Wu et al., 2021, Qiu et al., 8 Dec 2025).

The open directions identified by the papers are correspondingly diverse: alternative combinations of informativeness and cost beyond simple products or weighted sums, adaptive thresholds and confidence calibration, arbitrary region shapes beyond fixed squares, time-based rather than area-based cost models, stronger theory under noise or unknown supersets, recomputation of static region partitions during active learning, and closer integration of region-aware acquisition with annotation interfaces and workflow constraints (Colling et al., 2020, Kontonis et al., 2024, Hadjadj et al., 2023, Qiu et al., 8 Dec 2025).

Region-based active learning therefore occupies a broad but coherent research area. Its unifying principle is that active learning should operate at the level at which information and annotation cost are actually structured. In dense prediction that usually means local image or scene regions; in graph and nonparametric settings it means neighborhoods or region queries; in adaptive partitioning it means piecewise predictors over learned subdomains. The central technical problem is not merely how to rank uncertain samples, but how to define, score, and exploit regions so that each query captures informative local structure with minimal redundancy and realistic annotation cost.