DOTA-IOD: Incremental Detection in Aerial Images

• DOTA-IOD is a benchmark that reorganizes the DOTA dataset for incremental remote sensing object detection, highlighting challenges like catastrophic forgetting, scale variation, and unlabeled co-occurrence.
• It partitions 15 aerial classes into three sequential tasks where detectors learn without replaying previous data and only current-task annotations are provided.
• The framework employs response and topology distillation along with clustering-driven pseudo-labeling to mitigate forgetting and promote scale-aware learning.

Searching arXiv for DOTA-IOD and related benchmark papers to ground the article in current literature. DOTA-IOD is a remote-sensing incremental object detection benchmark introduced in “STAR-IOD: Scale-decoupled Topology Alignment with Pseudo-label Refinement for Remote Sensing Incremental Object Detection” (Zhang et al., 20 May 2026). In that work, DOTA-IOD denotes an incremental-learning version of the DOTA aerial object detection dataset, designed for rehearsal-free Remote Sensing Incremental Object Detection (RS-IOD). The benchmark is intended to evaluate detectors trained through sequential tasks in which each step introduces disjoint new classes, only current-task annotations are available, old-class instances may remain present but unlabeled, and previous-task training data are not revisited (Zhang et al., 20 May 2026). This places DOTA-IOD at the intersection of aerial-image detection, catastrophic forgetting, missing-annotation robustness, and scale-sensitive representation learning. The name can also be confused with broader “DOTA-based object detection” usage derived from the original DOTA benchmark (Xia et al., 2017), but in the strict literature-supported sense, “DOTA-IOD” refers specifically to the incremental benchmark defined in STAR-IOD (Zhang et al., 20 May 2026).

1. Origin in aerial object detection and the move to incremental learning

DOTA-IOD is derived from DOTA, the “Dataset for Object deTection in Aerial images,” which was introduced as a large-scale benchmark for object detection in aerial imagery and Earth Vision, also referred to as Earth Observation and Remote Sensing (Xia et al., 2017). The original DOTA paper motivated the benchmark by noting that progress in natural-scene detection had transferred poorly to aerial imagery because aerial objects exhibit huge variation in scale, arbitrary orientations, extreme aspect ratios, and dense arrangements, while available aerial datasets were limited in size, category diversity, complexity, and annotation detail (Xia et al., 2017).

The source DOTA dataset contains 2806 aerial images, 188,282 instances, and 15 categories, with images collected from different sensors and platforms and annotated by experts in aerial image interpretation (Xia et al., 2017). It uses arbitrary quadrilateral annotations with 8 degrees of freedom, reflecting the fact that horizontal bounding boxes often fit aerial objects poorly and can induce large overlaps in crowded scenes (Xia et al., 2017). DOTA was therefore constructed as a benchmark for object detection in aerial images rather than for incremental learning per se.

DOTA-IOD reinterprets this aerial detection setting under the RS-IOD regime introduced in STAR-IOD (Zhang et al., 20 May 2026). The paper argues that RS-IOD had lacked standardized remote-sensing incremental benchmarks and protocols, which made comparisons unfair or irreproducible, and that remote sensing adds distinctive complications relative to natural-image IOD, notably severe intra-class scale variation, complex backgrounds, dense layouts and occlusion, and missing old-class annotations during incremental learning (Zhang et al., 20 May 2026). DOTA-IOD was introduced to fill this benchmark gap.

A plausible implication is that DOTA-IOD should be understood not as a new underlying image collection, but as a task-specific reorganization of DOTA into a sequential-learning protocol emphasizing forgetting, old/new class coexistence, and annotation incompleteness (Zhang et al., 20 May 2026).

2. Formal RS-IOD setting instantiated by DOTA-IOD

STAR-IOD defines IOD and RS-IOD by partitioning the full class set into disjoint task subsets,

$$\mathcal{C}=\{\mathcal{C}_1,\mathcal{C}_2,\ldots,\mathcal{C}_n\}, \quad \mathcal{C}_i\cap \mathcal{C}_j=\emptyset \;\; \forall i\neq j.$$

At stage $t$, the detector is trained only on

$\mathcal{D}_t=\{X_t,Y_t\},$

where $Y_t$ contains annotations only for classes in $\mathcal{C}_t$ (Zhang et al., 20 May 2026). Images may still contain instances from all tasks, but old classes are unlabeled in new-task data, and the detector must learn new classes without access to previous-task datasets and without forgetting old ones (Zhang et al., 20 May 2026).

Within this formulation, DOTA-IOD is a concrete three-step multi-step incremental benchmark (Zhang et al., 20 May 2026). Training proceeds through sequential tasks, each introducing a disjoint set of categories. The previous model is frozen as a teacher, and the current model is trained as a student using only the current dataset, without rehearsal or replay of earlier datasets (Zhang et al., 20 May 2026). Old classes are handled through response distillation, topological distillation, and pseudo-label recovery (Zhang et al., 20 May 2026).

This setup differs materially from the original DOTA benchmark. DOTA defined two detection tasks, horizontal bounding box detection and oriented bounding box detection, with approximately 1/2 training, 1/6 validation, and 1/3 testing, and evaluated detectors using the same mAP calculation as PASCAL VOC (Xia et al., 2017). By contrast, DOTA-IOD is not primarily about orientation-sensitive localization; it is about sequential knowledge acquisition under incomplete labels, evaluated with standard COCO metrics and incremental old/new performance decompositions (Zhang et al., 20 May 2026).

This suggests that DOTA-IOD should be read as a benchmark transformation of DOTA from a static aerial detection task into a continual detection protocol.

3. Dataset construction, class sequence, and benchmark statistics

DOTA-IOD is built from DOTA, though the STAR-IOD paper does not state which DOTA version is used (Zhang et al., 20 May 2026). The benchmark contains 15 classes total, split into 3 tasks of 5 classes each (Zhang et al., 20 May 2026).

The task sequence is given explicitly in STAR-IOD.

Task	Classes
Task 1	Small-vehicle, Large-vehicle, Plane, Baseball-diamond, Ground-track-field
Task 2	Helicopter, Ship, Bridge, Soccer-ball-field, Tennis-court
Task 3	Storage-tank, Harbor, Roundabout, Basketball-court, Swimming-pool

The paper also reports train and test instance counts for each category (Zhang et al., 20 May 2026).

Category	Train instances	Test instances
Small-vehicle	49,864	10,833
Large-vehicle	36,243	9,257
Plane	16,426	5,298
Baseball-diamond	823	403
Ground-track-field	838	316
Helicopter	1,193	134
Ship	61,075	19,571
Bridge	3,634	896
Soccer-ball-field	849	351
Tennis-court	4,738	1,602
Storage-tank	9,639	5,025
Harbor	13,085	4,789
Roundabout	800	299
Basketball-court	1,119	297
Swimming-pool	3,324	763

Summing the table gives 204,650 total train instances and 59,834 total test instances (Zhang et al., 20 May 2026). The paper does not provide the total number of images, per-task image counts, or an explicit validation split (Zhang et al., 20 May 2026). It reports only train and test instance counts.

A notable property of DOTA-IOD is old/new co-occurrence. Among incremental-phase training images, defined as Only New + Co-occurrence, the co-occurrence rate is 57.4% for DOTA-IOD (Zhang et al., 20 May 2026). This is central to the benchmark’s difficulty: in more than half of incremental training images, new and old categories appear together, but only current-task classes are annotated. As a result, old-class instances are liable to be misclassified as background unless recovered through pseudo-labeling (Zhang et al., 20 May 2026).

In relation to the original DOTA dataset, this benchmark inherits the long-tailed and multi-scale character of aerial detection. DOTA itself reports that instance sizes are distributed as 57% small (10–50 pixels), 41% middle (50–300 pixels), and 2% large (above 300 pixels), and that some images contain up to 2000 instances with an average of 67.10 instances per image (Xia et al., 2017). Although STAR-IOD does not restate these statistics for DOTA-IOD, it explicitly frames scale variation and dense remote-sensing layouts as central RS-IOD difficulties (Zhang et al., 20 May 2026).

4. Annotation and evaluation conventions

The annotation regime in DOTA-IOD is indirect but important. Although the source DOTA dataset is an oriented aerial detection benchmark with arbitrary quadrilateral annotations (Xia et al., 2017), STAR-IOD conducts its experiments with Grounding DINO, uses standard DETR-style losses, and evaluates with standard COCO evaluation metrics:

$$mAP,\; mAP_{50},\; mAP_{75},\; mAP_s,\; mAP_m,\; mAP_l$$

(Zhang et al., 20 May 2026). In incremental evaluation, the paper further reports:

• $\mathbf{mAP^A}$ for all seen classes up to the current task,
• $\mathbf{mAP^P}$ for previous-task or old-class performance,
• $\mathbf{mAP^C}$ for current-task or newly introduced class performance (Zhang et al., 20 May 2026).

The STAR-IOD description strongly suggests that DOTA-IOD uses horizontal bounding boxes (HBB) rather than oriented-box evaluation (Zhang et al., 20 May 2026). The paper does not report rotated-box regression, angle parameterization, or oriented AP metrics, and its scale partitioning is based on standard bounding-box area thresholds,

$\tau_s=1024,\qquad \tau_m=9216$

(Zhang et al., 20 May 2026). It also does not specify the annotation conversion procedure from DOTA to the effective training format, nor whether original oriented boxes were converted to HBBs and how (Zhang et al., 20 May 2026).

This introduces a conceptual distinction between DOTA and DOTA-IOD. DOTA was expressly designed to benchmark both HBB and OBB detection, with oriented quadrilateral annotations as a central technical contribution (Xia et al., 2017). DOTA-IOD, by contrast, appears to operationalize DOTA under a COCO-style HBB incremental-detection pipeline (Zhang et al., 20 May 2026). The paper is explicit about the metrics but incomplete about preprocessing details such as image tiling, crop size, overlap, resizing pipeline, or exact annotation conversion (Zhang et al., 20 May 2026).

A plausible implication is that DOTA-IOD prioritizes a standardized continual-learning protocol over fidelity to DOTA’s original oriented-detection evaluation.

5. STAR-IOD framework and its relation to DOTA-IOD

DOTA-IOD is introduced together with STAR-IOD, the method proposed to address RS-IOD-specific difficulties (Zhang et al., 20 May 2026). The framework combines three elements: Subspace-decoupled Topology Distillation (STD), Clustering-driven Pseudo-label Generator (CPG), and Correspondence Response Distillation (CRD) (Zhang et al., 20 May 2026).

The motivation is twofold. First, remote-sensing objects exhibit strong intra-class scale variation, so transferring or preserving knowledge in a single undifferentiated feature space can undermine topology preservation (Zhang et al., 20 May 2026). Second, missing old-class annotations in new-task images cause old instances to be treated as background, which directly exacerbates forgetting (Zhang et al., 20 May 2026).

STD addresses scale-sensitive knowledge preservation by partitioning query features according to box area into three subspaces:

$t$0

with small, medium, and large categories defined by $t$1 and $t$2 (Zhang et al., 20 May 2026). For old classes present in a batch, it forms confidence-weighted, scale-specific prototypes,

$t$3

adds a background prototype from global image features,

$t$4

and computes inter-class relation matrices

$t$5

whose affinities are distilled from teacher to student through KL divergence (Zhang et al., 20 May 2026). The STD loss is

$t$6

CPG addresses missing old annotations by maintaining a per-class score bank $t$7 of teacher predictions above a minimum threshold $t$8, capped at bank size $t$9 (Zhang et al., 20 May 2026). The bank is split into low- and high-confidence clusters with K-means,

$\mathcal{D}_t=\{X_t,Y_t\},$0

and the class-specific adaptive threshold is defined as

$\mathcal{D}_t=\{X_t,Y_t\},$1

Teacher outputs above this threshold form an initial pseudo-label set,

$\mathcal{D}_t=\{X_t,Y_t\},$2

which is then de-duplicated against current-task ground truth using

$\mathcal{D}_t=\{X_t,Y_t\},$3

(Zhang et al., 20 May 2026).

CRD transfers teacher responses over old classes. Teacher probabilities are formed as

$\mathcal{D}_t=\{X_t,Y_t\},$4

and the classification alignment loss is

$\mathcal{D}_t=\{X_t,Y_t\},$5

with confidence-aware weight

$\mathcal{D}_t=\{X_t,Y_t\},$6

A regression distillation term is also applied:

$\mathcal{D}_t=\{X_t,Y_t\},$7

The total CRD is summed over decoder layers (Zhang et al., 20 May 2026).

The full STAR-IOD objective is

$\mathcal{D}_t=\{X_t,Y_t\},$8

with

$\mathcal{D}_t=\{X_t,Y_t\},$9

(Zhang et al., 20 May 2026). The framework uses Grounding DINO as the base detector, AdamW with initial learning rate $Y_t$0, weight decay $Y_t$1, 25 epochs per task, batch size 4, and NVIDIA L20 GPU hardware (Zhang et al., 20 May 2026).

6. Experimental performance, ablations, and failure modes

The core DOTA-IOD comparison in STAR-IOD evaluates Fine-tuning, MD-DETR, CL-DETR, ERD, GCD, and STAR-IOD under the multi-step incremental setting (Zhang et al., 20 May 2026). At Task 1, all Grounding-DINO-based methods share the same base-task performance:

• $Y_t$2
• $Y_t$3 (Zhang et al., 20 May 2026).

At Task 2, STAR-IOD reports:

• $Y_t$4
• $Y_t$5
• $Y_t$6
• $Y_t$7 which improves over GCD by +0.3, +0.5, +1.6, and +1.0, respectively (Zhang et al., 20 May 2026).

At Task 3, STAR-IOD reports:

• $Y_t$8
• $Y_t$9
• $\mathcal{C}_t$0
• $\mathcal{C}_t$1
• $\mathcal{C}_t$2
• $\mathcal{C}_t$3
• $\mathcal{C}_t$4
• $\mathcal{C}_t$5 (Zhang et al., 20 May 2026). Relative to GCD, the final-task gains are +2.1 on $\mathcal{C}_t$6, +1.2 on $\mathcal{C}_t$7, +3.2 on $\mathcal{C}_t$8, +1.4 on $\mathcal{C}_t$9, +1.8 on $$mAP,\; mAP_{50},\; mAP_{75},\; mAP_s,\; mAP_m,\; mAP_l$$0, +0.8 on $$mAP,\; mAP_{50},\; mAP_{75},\; mAP_s,\; mAP_m,\; mAP_l$$1, +3.2 on $$mAP,\; mAP_{50},\; mAP_{75},\; mAP_s,\; mAP_m,\; mAP_l$$2, and -0.3 on $$mAP,\; mAP_{50},\; mAP_{75},\; mAP_s,\; mAP_m,\; mAP_l$$3 (Zhang et al., 20 May 2026). The abstract’s claim of +2.1% mAP on DOTA-IOD refers to this final-task $$mAP,\; mAP_{50},\; mAP_{75},\; mAP_s,\; mAP_m,\; mAP_l$$4 improvement over the previous best method (Zhang et al., 20 May 2026).

The ablation on Scale-adaptive Instance Partitioning (SIP), which operationalizes the scale-decoupled aspect of STD, is reported directly on DOTA-IOD. Without SIP, the model obtains:

• $$mAP,\; mAP_{50},\; mAP_{75},\; mAP_s,\; mAP_m,\; mAP_l$$5
• $$mAP,\; mAP_{50},\; mAP_{75},\; mAP_s,\; mAP_m,\; mAP_l$$6
• $$mAP,\; mAP_{50},\; mAP_{75},\; mAP_s,\; mAP_m,\; mAP_l$$7
• $$mAP,\; mAP_{50},\; mAP_{75},\; mAP_s,\; mAP_m,\; mAP_l$$8
• $$mAP,\; mAP_{50},\; mAP_{75},\; mAP_s,\; mAP_m,\; mAP_l$$9 Whereas with SIP it achieves:
• $\mathbf{mAP^A}$0
• $\mathbf{mAP^A}$1
• $\mathbf{mAP^A}$2
• $\mathbf{mAP^A}$3
• $\mathbf{mAP^A}$4 (Zhang et al., 20 May 2026). The corresponding improvements are +2.1, +1.4, +0.7, +0.8, and +3.2 (Zhang et al., 20 May 2026). This directly supports the paper’s claim that scale-aware topology preservation is particularly beneficial on DOTA-IOD.

The paper also analyzes score bank size $\mathbf{mAP^A}$5 and concludes that performance improves consistently as $\mathbf{mAP^A}$6 increases from 5000 to 20000, that larger banks provide more reliable clustering statistics, and that gains beyond 20000 become marginal, though 40000 is slightly better; the chosen value is $\mathbf{mAP^A}$7 (Zhang et al., 20 May 2026).

Per-class final-stage DOTA-IOD results reveal large variation across categories (Zhang et al., 20 May 2026). Strong retained classes include:

• Tennis-court: AP 81.4, AP50 91.6
• Ship: AP 56.9, AP50 82.7
• Plane: AP 57.3, AP50 81.3 (Zhang et al., 20 May 2026). Newly introduced Task 3 classes remain learnable, for example:
• Storage-tank: AP50 72.5
• Harbor: AP50 76.9
• Basketball-court: AP50 63.5 (Zhang et al., 20 May 2026).

The major failure case highlighted by the authors is Helicopter. In incremental learning, Helicopter AP collapses to 0.9%, while the joint-training baseline for Helicopter AP is 36.7%; at Task 2 on DOTA-IOD, Helicopter AP50 is only 3.6%, far below 64.3% under joint training (Zhang et al., 20 May 2026). The authors attribute this to high semantic similarity between text embeddings of Plane and Helicopter, reliance on Grounding DINO text prompts, and the fact that helicopters are small and visually non-salient in aerial images (Zhang et al., 20 May 2026).

This failure is significant because it shows that even when incremental forgetting is mitigated at the benchmark level, semantic interference in vision-language detectors and the intrinsic small-object difficulty of aerial imagery remain unresolved.

7. Interpretation, scope, and unresolved ambiguities

DOTA-IOD is best understood as a standardized 15-class, 3-task, rehearsal-free RS-IOD benchmark derived from DOTA and evaluated with COCO-style horizontal-box metrics under severe old/new co-occurrence and missing annotation conditions (Zhang et al., 20 May 2026). Its immediate significance lies in providing a concrete evaluation protocol for catastrophic forgetting in remote sensing, where scale variation, dense object layouts, and unlabeled old classes are more pronounced than in many natural-image IOD settings (Zhang et al., 20 May 2026).

The benchmark’s relevance is heightened by its relation to the original DOTA dataset. DOTA had already established aerial imagery as a demanding detection domain because of tiny objects, arbitrary orientations, dense scenes, large aspect-ratio variation, and very large image sizes that require cropping and patch-based processing (Xia et al., 2017). DOTA-IOD transposes these detection challenges into a continual-learning regime, making preservation of old knowledge and recovery of missing old labels central evaluation criteria rather than secondary concerns (Zhang et al., 20 May 2026).

At the same time, the benchmark leaves several practical details unspecified. STAR-IOD does not clearly provide the exact DOTA version used, total image counts, an explicit validation split, the annotation conversion procedure, whether oriented boxes were converted to HBBs and how, the image patching or crop strategy, resizing details, the exact Grounding DINO variant, or the prompt format (Zhang et al., 20 May 2026). These omissions matter for exact reproducibility.

The term “DOTA-IOD” also admits broader informal interpretations in adjacent literature. In the original DOTA paper, the phrase would most naturally be read as object detection research on the DOTA benchmark for aerial images, even though the paper itself does not use the term “DOTA-IOD” literally (Xia et al., 2017). Outside remote sensing, similarly spelled names have referred to unrelated topics such as Dota 2 sequential item recommendation or the DOtA method for unsupervised 3D object detection from multi-agent LiDAR scans, but those are distinct uses and not the benchmark standardized in STAR-IOD (Dallmann et al., 2022, Xia et al., 11 Mar 2025).

In the strict encyclopedic sense supported by current arXiv literature, DOTA-IOD denotes the DOTA-derived benchmark for Remote Sensing Incremental Object Detection introduced by STAR-IOD (Zhang et al., 20 May 2026). It formalizes a three-stage, rehearsal-free continual-detection protocol over 15 aerial categories, emphasizes old/new co-occurrence and missing old annotations, and serves as an evaluation substrate for methods that combine distillation with pseudo-label refinement to mitigate catastrophic forgetting under strong intra-class scale variation.