---
title: 'STAR-IOD: Remote Sensing Incremental Detection'
url: https://www.emergentmind.com/topics/star-iod
type: topic
---

# STAR-IOD: Remote Sensing Incremental Detection

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" [2605.20738], 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 [2605.20738].

## 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 $\mathcal{D}_t = \{X_t, Y_t\}$ [2605.20738]. 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 [2605.20738]. 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 [2605.20738].

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 [2605.20738]. 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 [2605.20738]. 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 [2605.20738]. 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 $f_v$, a language backbone $f_l$, and a feature enhancer $f_e$ that fuses them as
$$
V, W = f_e(f_v(I), f_l(T))
$$
followed by a cross-modal decoder producing query embeddings $Q = \{q_i\}_{i=1}^{N}$ and predictions $\hat{y}_i = (\hat{s}_i, \hat{b}_i)$ [2605.20738]. The base detection objective is the standard DETR-style loss:
$$
\hat{\sigma} = \arg\min_{\sigma} \sum_{i=1}^{N} \mathcal{L}_{match}(\hat{y}_{\sigma(i)}, y_i)
$$
with
$$
\mathcal{L}_{match}(\hat{y}_{\sigma(i)}, y_i) = \mathcal{L}_{align}(\hat{s}_{\sigma(i)}, s_i) + \mathcal{L}_{reg}(\hat{b}_{\sigma(i)}, b_i)
$$
and
$$
\mathcal{L}_{detr}(\hat{y}, y) = \sum_{i=1}^{N} \big[ \mathcal{L}_{align}(\hat{s}_{\sigma(i)}, s_i) + \mathbb{1}_{\{s_i \neq \phi\}} \, \mathcal{L}_{reg}(\hat{b}_{\hat{\sigma}(i)}, b_i) \big]
$$
[2605.20738].

The incremental protocol is replay-free. In Task 1, Grounding DINO is trained on base classes using standard $\mathcal{L}_{detr}$, and the resulting model becomes the teacher for the next task [2605.20738]. At each incremental step $t \ge 2$, the previous teacher $\mathcal{M}_{t-1}$ is frozen, the student $\mathcal{M}_t$ is initialized from teacher weights, and training uses only the current dataset $\mathcal{D}_t$ [2605.20738]. 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 [2605.20738]. No replay buffer and no architecture expansion are introduced [2605.20738].

The total loss at each incremental step is
$$
\mathcal{L} = \mathcal{L}_{detr} + \lambda_{1}\mathcal{L}_{STD} + \mathcal{L}_{CRD}, \quad \lambda_1 = 3
$$
where $\mathcal{L}_{detr}$ is computed on real labels for new classes and pseudo-labels for old classes, $\mathcal{L}_{STD}$ performs topology alignment in scale-specific subspaces, and $\mathcal{L}_{CRD}$ distills per-query teacher responses [2605.20738]. 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 [2605.20738]. 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 [2605.20738].

The first stage is Scale-adaptive Instance Partitioning (SIP). Let $\mathcal{B} = \{ b_i \}_{i=1}^N$ be predicted boxes and $\mathcal{F} = \{ f_i \}_{i=1}^N$ their corresponding instance query features, with area $A(b_i)$ computed for each box [2605.20738]. Using thresholds $\tau_s$ and $\tau_m$, empirically set to $1024$ and $9216$ pixels, queries are split into three disjoint sets:
$$
\mathcal{Q}_{k} = \{f_i \mid \text{cond}_k(A(b_i))\},\quad k \in \mathcal{S} = \{s_s, s_m, s_l\}
$$
corresponding to small, medium, and large instances [2605.20738]. This decouples topology estimation across scales and prevents cross-scale interference.

Within each scale $k$ and old class $c \in \mathcal{C}_{1:(t-1)}$, STAR-IOD forms a confidence-weighted class prototype
$$
\mathbf{p}_{c}^{k} = \sum_{i \in \mathcal{Q}_k} \mathbb{1}(y_i = c) \cdot \frac{s_i}{\sum_{j} s_j} \cdot f_i
$$
where $y_i$ is the pseudo-label and $s_i$ is the classification confidence [2605.20738]. Only pseudo-labeled old-class queries contribute. In addition, a background prototype is defined from global image features $F_{img} \in \mathbb{R}^{H \times W \times D}$ as
$$
\mathbf{p}_{bg} = \frac{1}{H \times W} \sum_{x=1}^{H} \sum_{y=1}^{W} F_{img}(x,y)
$$
which acts as a background anchor in the topology graph [2605.20738].

Topology is then expressed through a relation matrix $\mathbf{M}^k \in \mathbb{R}^{N_k \times N_k}$:
$$
\mathbf{M}_{u,v}^k = \lVert \mathbf{p}_u^k - \mathbf{p}_v^k \rVert_2
$$
where $u,v$ index old classes present in that batch and scale [2605.20738]. Rather than directly matching distances, STAR-IOD converts them into semantic affinity distributions with a Boltzmann-type transform:
$$
\mathcal{P}_{u,v}^k = \frac{\exp(-M_{u,v}^k / \tau)}{\sum_{j=1}^{C'} \exp(-M_{u,j}^k / \tau)}
$$
with teacher and student each producing their own $\mathcal{P}_T^k$ and $\mathcal{P}_S^k$ [2605.20738]. The topology distillation loss is
$$
\mathcal{L}_{STD} = \sum_{k \in \mathcal{S}} \tau^2 \sum_{u=1}^{C'} \mathcal{D}_{KL}\left( \mathcal{P}_{T, u}^k \big\| \mathcal{P}_{S, u}^k \right)
$$
[2605.20738].

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 [2605.20738]. 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 [2605.20738]. 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 [2605.20738].

For each old class $c$, CPG maintains a score bank $\mathcal{Q}_c$ of teacher prediction scores [2605.20738]. From the current batch,
$$
S_{c} = \{ s_i \mid y_i=c, s_i > \delta_{min} \}
$$
with $\delta_{min} = 0.3$, and the bank is updated by concatenation:
$$
\mathcal{Q}_c = \text{Concat}(\mathcal{Q}_c, S_{c})
$$
subject to a maximum size $L_{cpg} = 20{,}000$ [2605.20738]. The method assumes that this bank tends to exhibit a bimodal structure, with low-confidence false positives separated from high-confidence true positives [2605.20738].

Thresholding is then recast as a 1-D, 2-cluster K-Means problem on $\mathcal{Q}_c$:
$$
\min_{\mu_{l}, \mu_{h}} \sum_{s \in \mathcal{C}_{l}} (s - \mu_{l})^2 + \sum_{s \in \mathcal{C}_{h}} (s - \mu_{h})^2
$$
where $\mathcal{C}_l$ and $\mathcal{C}_h$ are the low- and high-score clusters with centroids $\mu_l$ and $\mu_h$ [2605.20738]. The class-specific threshold is
$$
\tau_c = \min \{ s \mid s \in \mathcal{C}_{h} \}
$$
so all scores in the high-confidence cluster satisfy $s \ge \tau_c$ [2605.20738].

Pseudo-labels are generated from teacher detections $\mathcal{O} = \{(b_i, s_i, c_i)\}$ as
$$
\mathcal{P}_{r} = \{ (b_i, c_i) \mid s_i \geq \tau_{c_i} \}
$$
and then filtered against current-task ground truth $\mathcal{G}$:
$$
\mathcal{P} = \{ (b_p, c_p) \in \mathcal{P}_{r} \mid \max_{b_{gt} \in \mathcal{G}} \text{IoU}(b_p, b_{gt}) < \theta_{nms} \}
$$
with $\theta_{nms} = 0.7$ [2605.20738]. These pseudo-labels are added to $Y_t$ and used as targets for old classes in $\mathcal{L}_{detr}$ [2605.20738].

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 [2605.20738]. 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 [2605.20738].

## 5. Response distillation, datasets, and implementation

CRD, adopted from GCD, complements STD by distilling teacher outputs at the query level [2605.20738]. Teacher probabilities over old classes are
$$
P_i^{old} = \text{SoftMax}(\hat{s}_i^{old}/\tau)
$$
and student probabilities are computed similarly [2605.20738]. The classification distillation term is
$$
\mathcal{L}_{CRD}^{align} = \sum_{i=1}^{N} \alpha_i \mathcal{D}_{KL}(P_i^{old} \| P_i)
$$
where $\alpha_i = \max_{c \in \mathcal{C}_{1:t-1}} \hat{s}_i^{old}(c)$ weights each query by its maximum old-class probability [2605.20738]. Regression distillation is
$$
\mathcal{L}_{CRD}^{reg} = \sum_{i=1}^{N} \alpha_i \mathcal{L}_{bbox}(b_i^{old}, b_i)
$$
and the full CRD objective sums these losses over decoder layers:
$$
\mathcal{L}_{CRD} = \sum_{l=1}^{L} \left( \mathcal{L}_{CRD}^{align, l} + \mathcal{L}_{CRD}^{reg, l} \right)
$$
[2605.20738]. 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 [2605.20738]. DIOR-IOD is derived from DIOR and supports both a two-step 10+10 split and a multi-step 5+5+5+5 split [2605.20738]. 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 [2605.20738]. 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 [2605.20738].

The implementation uses Grounding DINO as the base architecture, AdamW with learning rate $1 \times 10^{-4}$ and weight decay $1 \times 10^{-4}$, 25 epochs per task, batch size 4, and one NVIDIA L20 GPU [2605.20738]. The STD scale thresholds are $(\tau_s, \tau_m) = (1024, 9216)$, and CPG uses $\delta_{min} = 0.3$, bank size $20{,}000$, and $\theta_{nms}=0.7$ [2605.20738]. 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 [2605.20738]. On DIOR-IOD in the two-step 10+10 setting at Task 2, fine-tuning yields $mAP^A = 22.6$ and $mAP^P = 1.3$, illustrating catastrophic forgetting, while GCD achieves $mAP^A = 46.9$ and $mAP^P = 39.4$ [2605.20738]. STAR-IOD improves this to $mAP^A = 48.6$ and $mAP^P = 40.7$, with gains also reported across scales: $mAP_s^A = 16.5$ versus 15.1, $mAP_m^A = 41.4$ versus 39.3, and $mAP_l^A = 69.0$ versus 66.9 [2605.20738]. In the DIOR-IOD multi-step 5+5+5+5 setting at Task 4, STAR-IOD reaches $mAP^A = 36.9$ and $mAP^P = 37.7$, compared with GCD’s $34.5$ and $33.6$ [2605.20738]. On DOTA-IOD at Task 3, STAR-IOD reports $mAP^A = 39.4$, $mAP_{75}^A = 44.1$, and $mAP^P = 42.7$, compared with GCD’s $37.3$, $40.9$, and $41.3$ [2605.20738].

Ablation studies clarify the functional role of each module. On DIOR-IOD two-step, CRD alone raises $mAP^A$ from 22.6 to 31.2 and $mAP^P$ from 1.5 to 8.3; CPG alone reaches $mAP^A = 43.5$ and $mAP^P = 30.4$; CRD plus CPG achieves $mAP^A = 46.4$ and $mAP^P = 40.7$; and adding STD produces the full model with $mAP^A = 48.6$ and $mAP^C = 57.4$ [2605.20738]. 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 [2605.20738]. In a separate ablation on DOTA-IOD, removing SIP and using a single global topology matrix reduces $mAP^A$ from 39.4 to 37.3 and $mAP_l$ from 48.4 to 45.2 [2605.20738], 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 $\delta_{min} = 0.3$ yields $mAP^A = 46.5$, $mAP^P = 41.1$, and $mAP^P_{50} = 65.6$; a best-case static threshold of $\delta = 0.4$ yields $mAP^A = 45.3$ and $mAP^P = 39.7$; Consistent Teacher gives $46.1$, $40.2$, and $64.8$; and LPL gives $45.5$, $40.0$, and $61.9$ [2605.20738]. 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 [2605.20738]. 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 [2605.20738]. 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 [2605.20738].

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 $+1.7$ mAP and $+2.1$ 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 [2605.20738].

Source: https://www.emergentmind.com/topics/star-iod