STAR-IOD: Remote Sensing Incremental Detection

• STAR-IOD is a remote sensing incremental object detection framework that addresses catastrophic forgetting and scale variation using Subspace-decoupled Topology Distillation.
• It augments a DETR-style detector with Clustering-driven Pseudo-label Generation and Correspondence Response Distillation to handle missing annotations and stabilize predictions.
• Empirical evaluations on DIOR-IOD and DOTA-IOD benchmarks demonstrate notable mAP improvements over previous methods, all without replaying historical images.

Searching arXiv for STAR-IOD and closely related incremental detection work. STAR-IOD is a framework for Remote Sensing Incremental Object Detection (RS-IOD) that addresses two coupled failure modes of continual detection in aerial imagery: catastrophic forgetting of previously learned categories and degradation caused by intra-class scale variation and missing annotations for old classes. Introduced in "STAR-IOD: Scale-decoupled Topology Alignment with Pseudo-label Refinement for Remote Sensing Incremental Object Detection" (Zhang et al., 20 May 2026), it is built on Grounding DINO and augments standard DETR-style training with three components: Subspace-decoupled Topology Distillation (STD), Clustering-driven Pseudo-label Generator (CPG), and Correspondence Response Distillation (CRD). Within the RS-IOD setting, STAR-IOD is designed to preserve old-class knowledge while learning new classes from streaming remote sensing data without replaying historical images, and it is evaluated on two benchmarks constructed for this purpose, DIOR-IOD and DOTA-IOD (Zhang et al., 20 May 2026).

1. Problem formulation and scope

RS-IOD is defined by a class set $$\mathcal{C} = \{\mathcal{C}_1, \dots, \mathcal{C}_n\}$$ split into disjoint subsets per stage, with training proceeding in stages $t=1,\dots,n$ and the detector updated from $\mathcal{M}_{t-1}$ to $\mathcal{M}_t$ using only the current dataset $\mathcal{D}_t = \{X_t, Y_t\}$ (Zhang et al., 20 May 2026). At stage $t$, only classes in $\mathcal{C}_t$ are annotated, although images may contain objects from any class in $\mathcal{C}$; consequently, old-class objects are unlabeled and treated as background (Zhang et al., 20 May 2026). The objective is to maintain performance on all past classes $\mathcal{C}_{1:(t-1)}$ while learning new classes $\mathcal{C}_t$, without access to past images (Zhang et al., 20 May 2026).

In remote sensing imagery, this setting is aggravated by scene properties that differ materially from natural-image incremental detection. The same category can appear at widely different physical and pixel scales, and features of a single class may cluster by scale more strongly than by semantic identity (Zhang et al., 20 May 2026). This undermines prototype-based or globally distilled representations, because a single class prototype can average over several scale-dependent modes rather than preserving them separately. A second complication is pervasive missing annotation for old classes during later tasks: in the incremental-phase training data, co-occurrence of old and new classes is common, with old instances left unlabeled and thus penalized as false positives if predicted (Zhang et al., 20 May 2026). STAR-IOD is organized around these two issues rather than treating them as secondary nuisances.

This focus distinguishes STAR-IOD from generic incremental object detection methods that primarily rely on response distillation or fixed-threshold pseudo-labeling. The paper characterizes prior methods such as LwF-derived distillation, CL-DETR, MD-DETR, ERD, and GCD as typically treating each class as a homogeneous feature set, overlooking scale-dependent structure, and using fixed thresholds or simple heuristics for pseudo-labeling (Zhang et al., 20 May 2026). A plausible implication is that STAR-IOD should be read not merely as another replay-free continual detector, but as a remote-sensing-specific reformulation of what must be preserved during incremental learning.

2. Architectural foundation and training protocol

STAR-IOD is built on Grounding DINO, described in the paper as a vision-language Transformer-based detector with a visual backbone $t=1,\dots,n$0, a language backbone $t=1,\dots,n$1, and a feature enhancer $t=1,\dots,n$2 that fuses them as

$t=1,\dots,n$3

followed by a cross-modal decoder producing query embeddings $t=1,\dots,n$4 and predictions $t=1,\dots,n$5 (Zhang et al., 20 May 2026). The base detection objective is the standard DETR-style loss:

$t=1,\dots,n$6

with

$t=1,\dots,n$7

and

$t=1,\dots,n$8

(Zhang et al., 20 May 2026).

The incremental protocol is replay-free. In Task 1, Grounding DINO is trained on base classes using standard $t=1,\dots,n$9, and the resulting model becomes the teacher for the next task (Zhang et al., 20 May 2026). At each incremental step $\mathcal{M}_{t-1}$0, the previous teacher $\mathcal{M}_{t-1}$1 is frozen, the student $\mathcal{M}_{t-1}$2 is initialized from teacher weights, and training uses only the current dataset $\mathcal{M}_{t-1}$3 (Zhang et al., 20 May 2026). The teacher is run on current-task images to provide pseudo-labels for old classes through CPG, query-level predictions for CRD, and features and prototypes for STD (Zhang et al., 20 May 2026). No replay buffer and no architecture expansion are introduced (Zhang et al., 20 May 2026).

The total loss at each incremental step is

$\mathcal{M}_{t-1}$4

where $\mathcal{M}_{t-1}$5 is computed on real labels for new classes and pseudo-labels for old classes, $\mathcal{M}_{t-1}$6 performs topology alignment in scale-specific subspaces, and $\mathcal{M}_{t-1}$7 distills per-query teacher responses (Zhang et al., 20 May 2026). This decomposition is central to the method’s logic: pseudo-labeling repairs missing supervision, topology distillation preserves structural organization among old classes, and response distillation stabilizes query-wise outputs.

3. Subspace-decoupled Topology Distillation

STD is the defining mechanism of STAR-IOD. Its premise is that “structural knowledge” should be interpreted as the inter-class topological relationships in feature space, rather than only logits or per-instance embeddings (Zhang et al., 20 May 2026). In remote sensing, those relationships are corrupted when all instances of a class are pooled into a single cluster despite strong internal scale variation. STD addresses this by partitioning instance features into scale-specific subspaces and aligning teacher-student class topology within each subspace separately (Zhang et al., 20 May 2026).

The first stage is Scale-adaptive Instance Partitioning (SIP). Let $\mathcal{M}_{t-1}$8 be predicted boxes and $\mathcal{M}_{t-1}$9 their corresponding instance query features, with area $\mathcal{M}_t$0 computed for each box (Zhang et al., 20 May 2026). Using thresholds $\mathcal{M}_t$1 and $\mathcal{M}_t$2, empirically set to $\mathcal{M}_t$3 and $\mathcal{M}_t$4 pixels, queries are split into three disjoint sets:

$\mathcal{M}_t$5

corresponding to small, medium, and large instances (Zhang et al., 20 May 2026). This decouples topology estimation across scales and prevents cross-scale interference.

Within each scale $\mathcal{M}_t$6 and old class $\mathcal{M}_t$7, STAR-IOD forms a confidence-weighted class prototype

$\mathcal{M}_t$8

where $\mathcal{M}_t$9 is the pseudo-label and $\mathcal{D}_t = \{X_t, Y_t\}$0 is the classification confidence (Zhang et al., 20 May 2026). Only pseudo-labeled old-class queries contribute. In addition, a background prototype is defined from global image features $\mathcal{D}_t = \{X_t, Y_t\}$1 as

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

which acts as a background anchor in the topology graph (Zhang et al., 20 May 2026).

Topology is then expressed through a relation matrix $\mathcal{D}_t = \{X_t, Y_t\}$3:

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

where $\mathcal{D}_t = \{X_t, Y_t\}$5 index old classes present in that batch and scale (Zhang et al., 20 May 2026). Rather than directly matching distances, STAR-IOD converts them into semantic affinity distributions with a Boltzmann-type transform:

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

with teacher and student each producing their own $\mathcal{D}_t = \{X_t, Y_t\}$7 and $\mathcal{D}_t = \{X_t, Y_t\}$8 (Zhang et al., 20 May 2026). The topology distillation loss is

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

(Zhang et al., 20 May 2026).

The significance of STD lies in what it preserves. It does not force the student to reproduce absolute feature coordinates; instead it preserves teacher-like inter-class relationships within small-, medium-, and large-object subspaces (Zhang et al., 20 May 2026). This suggests a more geometric conception of continual learning in remote sensing, where retention depends on maintaining the arrangement of class manifolds under scale-conditioned feature statistics rather than only matching outputs.

4. Clustering-driven Pseudo-label Generator

CPG addresses the missing-annotation problem that arises when old-class instances appear in incremental-step images but are unlabeled. In the constructed benchmarks, co-occurrence of old and new classes is frequent: 23.5% of incremental-phase training images in DIOR-IOD and 57.4% in DOTA-IOD contain both old and new classes with old objects unlabeled (Zhang et al., 20 May 2026). Naive training would therefore suppress old-class detections by penalizing them as background. Pseudo-labeling is the obvious countermeasure, but fixed thresholds are brittle because score distributions vary by class and shift across tasks (Zhang et al., 20 May 2026).

For each old class $t$0, CPG maintains a score bank $t$1 of teacher prediction scores (Zhang et al., 20 May 2026). From the current batch,

$t$2

with $t$3, and the bank is updated by concatenation:

$t$4

subject to a maximum size $t$5 (Zhang et al., 20 May 2026). The method assumes that this bank tends to exhibit a bimodal structure, with low-confidence false positives separated from high-confidence true positives (Zhang et al., 20 May 2026).

Thresholding is then recast as a 1-D, 2-cluster K-Means problem on $t$6:

$t$7

where $t$8 and $t$9 are the low- and high-score clusters with centroids $\mathcal{C}_t$0 and $\mathcal{C}_t$1 (Zhang et al., 20 May 2026). The class-specific threshold is

$\mathcal{C}_t$2

so all scores in the high-confidence cluster satisfy $\mathcal{C}_t$3 (Zhang et al., 20 May 2026).

Pseudo-labels are generated from teacher detections $\mathcal{C}_t$4 as

$\mathcal{C}_t$5

and then filtered against current-task ground truth $\mathcal{C}_t$6:

$\mathcal{C}_t$7

with $\mathcal{C}_t$8 (Zhang et al., 20 May 2026). These pseudo-labels are added to $\mathcal{C}_t$9 and used as targets for old classes in $\mathcal{C}$0 (Zhang et al., 20 May 2026).

The method’s rationale is explicit: class-wise adaptive thresholds are better suited than fixed global thresholds in dense remote sensing scenes, where clutter, occlusion, and class-dependent score distributions make uniform thresholding unstable (Zhang et al., 20 May 2026). The paper further states that K-Means is preferred because it makes no parametric distribution assumption, unlike Gaussian mixture-based schemes such as Consistent Teacher (Zhang et al., 20 May 2026).

5. Response distillation, datasets, and implementation

CRD, adopted from GCD, complements STD by distilling teacher outputs at the query level (Zhang et al., 20 May 2026). Teacher probabilities over old classes are

$\mathcal{C}$1

and student probabilities are computed similarly (Zhang et al., 20 May 2026). The classification distillation term is

$\mathcal{C}$2

where $\mathcal{C}$3 weights each query by its maximum old-class probability (Zhang et al., 20 May 2026). Regression distillation is

$\mathcal{C}$4

and the full CRD objective sums these losses over decoder layers:

$\mathcal{C}$5

(Zhang et al., 20 May 2026). In the overall design, CRD stabilizes per-query responses, while STD preserves broader class geometry.

To support RS-IOD evaluation, the paper constructs two datasets from established remote-sensing detection corpora (Zhang et al., 20 May 2026). DIOR-IOD is derived from DIOR and supports both a two-step 10+10 split and a multi-step 5+5+5+5 split (Zhang et al., 20 May 2026). In the two-step split, Task 1 contains airplane, airport, bridge, service-area, toll-station, harbor, overpass, ship, trainstation, and vehicle, while Task 2 contains baseballfield, basketballcourt, chimney, dam, golffield, groundtrackfield, stadium, storagetank, tenniscourt, and windmill (Zhang et al., 20 May 2026). DOTA-IOD is derived from DOTA and uses three 5-class tasks: Task 1 includes small-vehicle, large-vehicle, plane, baseball-diamond, and ground-track-field; Task 2 includes helicopter, ship, bridge, soccer-ball-field, and tennis-court; Task 3 includes storage-tank, harbor, roundabout, basketball-court, and swimming-pool (Zhang et al., 20 May 2026).

The implementation uses Grounding DINO as the base architecture, AdamW with learning rate $\mathcal{C}$6 and weight decay $\mathcal{C}$7, 25 epochs per task, batch size 4, and one NVIDIA L20 GPU (Zhang et al., 20 May 2026). The STD scale thresholds are $\mathcal{C}$8, and CPG uses $\mathcal{C}$9, bank size $\mathcal{C}_{1:(t-1)}$0, and $\mathcal{C}_{1:(t-1)}$1 (Zhang et al., 20 May 2026). These details matter because the paper frames STAR-IOD as replay-free and plug-and-play, not as a system requiring memory buffers or network expansion.

6. Empirical results, interpretation, and limitations

The reported results show consistent gains over prior incremental detectors on both benchmarks (Zhang et al., 20 May 2026). On DIOR-IOD in the two-step 10+10 setting at Task 2, fine-tuning yields $\mathcal{C}_{1:(t-1)}$2 and $\mathcal{C}_{1:(t-1)}$3, illustrating catastrophic forgetting, while GCD achieves $\mathcal{C}_{1:(t-1)}$4 and $\mathcal{C}_{1:(t-1)}$5 (Zhang et al., 20 May 2026). STAR-IOD improves this to $\mathcal{C}_{1:(t-1)}$6 and $\mathcal{C}_{1:(t-1)}$7, with gains also reported across scales: $\mathcal{C}_{1:(t-1)}$8 versus 15.1, $\mathcal{C}_{1:(t-1)}$9 versus 39.3, and $\mathcal{C}_t$0 versus 66.9 (Zhang et al., 20 May 2026). In the DIOR-IOD multi-step 5+5+5+5 setting at Task 4, STAR-IOD reaches $\mathcal{C}_t$1 and $\mathcal{C}_t$2, compared with GCD’s $\mathcal{C}_t$3 and $\mathcal{C}_t$4 (Zhang et al., 20 May 2026). On DOTA-IOD at Task 3, STAR-IOD reports $\mathcal{C}_t$5, $\mathcal{C}_t$6, and $\mathcal{C}_t$7, compared with GCD’s $\mathcal{C}_t$8, $\mathcal{C}_t$9, and $t=1,\dots,n$00 (Zhang et al., 20 May 2026).

Ablation studies clarify the functional role of each module. On DIOR-IOD two-step, CRD alone raises $t=1,\dots,n$01 from 22.6 to 31.2 and $t=1,\dots,n$02 from 1.5 to 8.3; CPG alone reaches $t=1,\dots,n$03 and $t=1,\dots,n$04; CRD plus CPG achieves $t=1,\dots,n$05 and $t=1,\dots,n$06; and adding STD produces the full model with $t=1,\dots,n$07 and $t=1,\dots,n$08 (Zhang et al., 20 May 2026). The paper interprets this as evidence that CPG is crucial for retaining old classes, CRD further stabilizes both old and new classes, and STD recovers new-class plasticity lost when preservation constraints become strong (Zhang et al., 20 May 2026). In a separate ablation on DOTA-IOD, removing SIP and using a single global topology matrix reduces $t=1,\dots,n$09 from 39.4 to 37.3 and $t=1,\dots,n$10 from 48.4 to 45.2 (Zhang et al., 20 May 2026), supporting the claim that scale-aware topology is essential rather than incidental.

The paper also includes a comparative analysis of pseudo-labeling strategies. On DIOR-IOD, the best CPG configuration with $t=1,\dots,n$11 yields $t=1,\dots,n$12, $t=1,\dots,n$13, and $t=1,\dots,n$14; a best-case static threshold of $t=1,\dots,n$15 yields $t=1,\dots,n$16 and $t=1,\dots,n$17; Consistent Teacher gives $t=1,\dots,n$18, $t=1,\dots,n$19, and $t=1,\dots,n$20; and LPL gives $t=1,\dots,n$21, $t=1,\dots,n$22, and $t=1,\dots,n$23 (Zhang et al., 20 May 2026). This supports the paper’s claim that simple class-wise K-Means thresholding is more robust than either static thresholds or GMM-based adaptive alternatives in the RS-IOD setting.

The principal limitation highlighted in the paper concerns semantically similar classes in Grounding DINO’s multimodal space. Helicopter on DOTA-IOD is identified as a notable failure case, with AP collapsing from 36.7% in joint training to 0.9% in incremental RS-IOD (Zhang et al., 20 May 2026). The analysis attributes this to high similarity between the text embeddings for “plane” and “helicopter” and to weak visual cues for small, low-resolution helicopters (Zhang et al., 20 May 2026). This suggests that STAR-IOD’s topology and pseudo-label mechanisms do not fully solve semantic ambiguity induced by the underlying vision-language representation. The authors therefore point to future directions involving more discriminative text encoders, stronger foundation models, and extensions of STD and CPG to other continual tasks and other multi-scale domains (Zhang et al., 20 May 2026).

Taken together, STAR-IOD can be understood as a replay-free continual detection framework whose central innovation is to redefine knowledge preservation in remote sensing as scale-conditioned topology preservation plus adaptive recovery of old-class supervision. Its empirical gains on DIOR-IOD and DOTA-IOD, reported as $t=1,\dots,n$24 mAP and $t=1,\dots,n$25 mAP over previous state of the art on the main settings, respectively, follow directly from that design logic rather than from architectural expansion or exemplar storage (Zhang et al., 20 May 2026).