LOGOS: Language-guided Oriented Object Detection in Aerial Scenes

Abstract: Object detection in geospatial scenes, such as satellite and aerial imagery, poses significant challenges due to the varying orientations and densities of objects, as well as the complex backgrounds inherent to remote sensing imagery. Traditional methods for oriented object detection have struggled to address issues such as angular discontinuity, fixed query sizes, and inefficiencies in handling sparse or cluttered scenes. In this paper, we propose LOGOS, a novel transformer-based approach that leverages textual prompts to guide the detection of oriented objects in aerial scenes. In particular, our proposed approach incorporates prompt-modulated content queries to dynamically adjust the model's focus based on the provided text, thereby improving object detection accuracy in complex environments. Empirically, extensive experiments on the DOTA dataset demonstrate that LOGOS outperforms existing state-of-the-art methods, particularly in densely packed and rotated object scenarios. Our approach offers a significant step forward in improving the robustness and scalability of oriented object detection in remote sensing applications.

• The paper introduces LOGOS, a DINO-based detector that uses FiLM-modulated queries, visual-text cross-attention, prompt-based class masking, and sine/cosine angle encoding to address orientation ambiguity and clutter in aerial imagery.
• The paper achieves state-of-the-art results on DOTA-v1.0, v1.5, and v2.0, reaching 81.32, 69.97, and 66.04 mAP respectively, with its largest improvement—8.38 points—on DOTA-v2.0.
• The paper shows that language conditioning is particularly useful for long-tailed and cluttered scenes, but performance remains weaker for ships and square-like objects while evaluation still relies on accurate category prompts.

Motivation and problem setting

Oriented object detection (OBB prediction) in optical remote sensing imagery must contend with arbitrary object orientations, extreme scale variation, dense instance packing, and heavy background clutter. The authors identify three persistent failure modes in prior work: (i) angular periodicity, where regression- or classification-based angle heads produce inconsistent predictions at the $-\pi/2$–$\pi/2$ boundary; (ii) orientation ambiguity for square-like objects; and (iii) fixed query counts in DETR-style detectors, which waste capacity on sparse scenes and underfit densely packed ones. LOGOS addresses these by conditioning a DETR-style detector on a textual prompt that enumerates the categories present in an image, so that both query initialization and cross-attention are modulated by language. The task is formalized as predicting a set of oriented boxes $\mathbf{b} = (x, y, w, h, \theta)$ with $\theta \in (-\pi, \pi]$, factorized per-query as $p(\mathbf{d}_q \mid \mathbf{I}, \mathbf{P})$.

Architecture

LOGOS follows an encoder–decoder design built on DINO (Zhang et al., 2022). The encoder produces multi-scale visual tokens $$\mathbf{enc}(\mathbf{I}) \in \mathbb{R}^{M \times D}$$, while the prompt is embedded as $$\mathbf{emb}(\mathbf{P}) \in \mathbb{R}^{K \times D}$$. Two components carry the language conditioning:

• Prompt-modulated content queries: learnable content queries are transformed via Feature-wise Linear Modulation (FiLM) using a pooled prompt representation, so the initial query set is semantically conditioned rather than category-agnostic.
• Text-aware cross-attention: decoder queries attend over the concatenation of visual and text tokens as keys and values, allowing each query to draw on both spatial evidence and prompt semantics.

The prediction head regresses $(x, y, w, h, \sin\theta, \cos\theta)$, encoding the angle with sine/cosine pairs to avoid boundary discontinuities — a direct response to the angular-periodicity critique of Oriented R-CNN and $R^3$Det. A class mask derived from the prompt suppresses logits for classes outside the prompted set $S$, filtering irrelevant categories at both training and inference.

Training uses Hungarian matching with a cost combining focal classification loss, L1 box loss, and rotated GIoU, plus a contrastive denoising (CDN) auxiliary loss following DN-DETR (Li et al., 2022) on positive and negative noised queries. Inference applies a relevance-score threshold followed by oriented NMS. Notably, despite the DETR lineage's end-to-end set-prediction motivation, NMS is retained at inference — a pragmatic but somewhat inconsistent choice relative to the stated advantages of DETR-style matching.

Experimental results

Evaluation covers DOTA-v1.0, v1.5, and v2.0 with mAP@0.5:0.95, using a ResNet-50 backbone, 900 decoder queries (300 × 3 patterns, matched to DN-DETR's compute budget), AdamW, and a single A100.

Dataset	LOGOS mAP	Best prior	Margin
DOTA-v1.0	81.32	80.87 (Oriented R-CNN)	+0.45
DOTA-v1.5	69.97	69.56 (RoI Transformer)	+0.41
DOTA-v2.0	66.04	57.66 (DCFL-R101)	+8.38

On DOTA-v1.0, LOGOS attains the best per-class AP on Plane (95.23), Storage Tank (93.81), Small Boat Facility (74.54), Roundabout (70.86), and Harbor (93.50) — the Harbor result exceeding the next-best method by more than 11 points. On v1.5 it leads on Plane (93.24), Bridge (52.58), Large Vehicle (82.00), Traffic Cone (94.60), Harbor (91.98), Swimming Pool (80.15), and Container Crane (60.24); the TC and CC gains are substantial given how poorly most baselines handle these rare classes. The largest aggregate gain appears on DOTA-v2.0 (+8.38 mAP over DCFL-R101), where LOGOS takes first place on 13 of 18 classes, including Airplane (89.15) and Helicopter (67.71). These results support the claim that prompt conditioning is most valuable when class distributions are long-tailed or scenes are cluttered.

Two caveats temper the headline numbers. First, margins on v1.0 and v1.5 are modest (under half a point), so the improvement is not uniformly decisive across benchmark versions. Second, LOGOS underperforms on Ship (65.58 on v1.5 versus ~80+ for several baselines) and Baseball Diamond, which the authors attribute to scale and prompt-conditioning limitations. Qualitative figures confirm strong behavior in harbors and industrial scenes but also show failures under very high small-vehicle density and for isolated objects at unusual orientations.

Limitations and open questions

The paper concedes several points explicitly. Performance degrades on Ship and Baseball Diamond across all three benchmarks, suggesting the FiLM-based pooling may discard fine-grained semantic distinctions relevant to those classes. The conclusion notes residual difficulty near 0°/180° orientations in dense scenes, indicating the sine/cosine parameterization does not fully eliminate boundary sensitivity. The evaluation protocol assumes oracle-like prompts — the prompt lists exactly the ground-truth categories present — so robustness to noisy, partial, or open-vocabulary prompts is untested. Ablations isolating the contribution of FiLM modulation versus text-aware cross-attention versus the class mask are also absent, leaving open which component drives the v2.0 gain. Finally, real-time applicability remains unaddressed.

Conclusion

LOGOS demonstrates that injecting textual prompts into a DINO-based oriented detector — through FiLM-modulated content queries, joint visual-text cross-attention, and prompt-derived class masking — yields state-of-the-art mAP on all three DOTA versions, with the largest margin on the hardest v2.0 split. The consistent weakness on Ship and square-like categories, together with the reliance on ground-truth-informed prompts, defines the clearest open questions for follow-up work on language-conditioned oriented detection.