💻 computer science

LOGOS: Language-guided Oriented Object Detection in Aerial Scenes

The paper proposes LOGOS, a novel transformer-based method that leverages textual prompts to guide oriented object detection in aerial scenes, effectively addressing challenges like angular discontinuity and complex backgrounds while outperforming state-of-the-art approaches on the DOTA dataset.

Original authors: Trong-Thuan Nguyen, Minh-Triet Tran

Published 2026-07-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Trong-Thuan Nguyen, Minh-Triet Tran

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are looking at a giant, high-resolution photo taken from a drone or a satellite. It's a bird's-eye view of a busy city, a harbor full of ships, or an airfield packed with planes. Now, try to find every single object in that picture. The problem is, these aren't just boxes sitting neatly on a shelf; they are scattered everywhere, tilted at weird angles, and often squished together in messy piles.

For a long time, computer programs trying to do this were like a clumsy librarian who only knows how to look for books in perfect, straight rows. If a book was lying diagonally or if the shelf was crowded, the librarian got confused. They would try to guess the angle of every single object, often getting stuck in a loop where they couldn't tell if a car was pointing left or right because the math got messy. They also had a fixed number of "searchers" (called queries) they could send out. If there were too many objects, the searchers got overwhelmed; if there were too few, they wasted time looking for ghosts.

Enter LOGOS, a new approach proposed by researchers Trong-Thuan Nguyen and Minh-Triet Tran. Think of LOGOS as giving that librarian a very specific, helpful note before they start searching. Instead of just saying, "Find everything," you hand them a note that says, "Find the ships in the harbor" or "Look for planes at the airport."

Here is how it works in plain terms:

The Magic Note: The system takes a text prompt (like "find cars") and uses it to tune its "searchers." It's like giving the searchers a pair of glasses that only lets them see the specific things you asked for.
Smart Searching: Instead of blindly scanning the whole image with a fixed number of searchers, LOGOS uses these text-guided searchers to focus their attention exactly where it matters. If you ask for "ships," the model ignores the buildings and the roads, saving energy and getting better results.
No More Confusion: The old methods struggled with the math of rotation (like getting confused between 90 degrees and -90 degrees). LOGOS handles this by encoding the angles in a way that keeps the math smooth, so it doesn't trip over itself when objects are tilted.

The researchers tested this on the DOTA dataset, a massive collection of aerial images with over 280,000 labeled objects. They compared LOGOS against the current best methods (the "state-of-the-art").

The Results:
The paper shows that LOGOS is a strong contender. On the DOTA-v1.0 dataset, LOGOS achieved a score (called mAP) of 81.32%, beating many other top methods. It did particularly well at finding planes, storage tanks, and harbors. For example, it found 93.50% of the roundabouts and 93.81% of the storage tanks.

On DOTA-v1.5, it scored 69.97%, and on the newer DOTA-v2.0, it hit 66.04%. In these tests, the text prompts helped the model filter out noise. For instance, when asked to find "traffic cones," the model scored a massive 94.60%, showing that the text guidance really helps it focus on small, specific things.

However, the paper is honest about where it still stumbles. The model had a bit more trouble with ships and baseball courts compared to other categories. The authors suggest this might be because these objects are tricky to spot even with the text help, or perhaps they need more specific training. They also point out that in extremely crowded scenes, like a harbor with ships packed shoulder-to-shoulder, the model sometimes misses objects or gets confused, just like a human would in a chaotic crowd.

The researchers don't claim this is the final, perfect solution to all aerial detection problems. Instead, they suggest that by using text to guide the search, they have taken a significant step forward in making these systems more robust and scalable. They believe this approach could be useful for things like urban planning and disaster management, but they also note that more work is needed to handle extreme angles and to make the system fast enough for real-time use.

In short, LOGOS is like giving a computer a specific instruction manual before it starts a search, helping it ignore the clutter and find exactly what you're looking for, even when the world is tilted and messy. It's not perfect yet, but it's a clever new way to tackle a very difficult puzzle.

Technical Summary: LOGOS – Language-guided Oriented Object Detection in Aerial Scenes

Problem Statement

Object detection in optical remote sensing imagery (satellite and aerial) presents unique challenges distinct from natural scene detection. Objects such as buildings, vehicles, and ships appear at arbitrary orientations, necessitating Oriented Bounding Boxes (OBBs) rather than axis-aligned boxes. Existing methods face three primary limitations:

Angular Discontinuity: Traditional angle prediction struggles with periodicity (e.g., $-\pi/2$ and $\pi/2$ being equivalent), leading to inconsistent predictions.
Fixed Query Inefficiency: Transformer-based models (e.g., DETR) typically use a fixed number of object queries. This creates inefficiencies in remote sensing scenes where object density varies wildly, leading to either excessive computational overhead in sparse scenes or insufficient coverage in dense, cluttered environments.
Orientation Ambiguity: Square-like bounding boxes and fixed angular ranges often cause confusion during training, particularly for objects appearing at angles outside predefined ranges.

Furthermore, current methods often lack mechanisms to incorporate specific semantic guidance, making it difficult to filter relevant objects in highly cluttered backgrounds.

Methodology: LOGOS

The authors propose LOGOS (Language-guided Oriented Object Detection in Aerial Scenes), a novel transformer-based framework that conditions the detection process on textual prompts. The architecture follows an encoder-decoder structure with the following key components:

1. Prompt-Modulated Content Queries

Instead of using static, learnable queries, LOGOS dynamically adjusts the initial content queries based on the input text prompt.

Mechanism: The model employs Feature-wise Linear Modulation (FiLM). The learnable content queries ( $\mathbf{q}^{(0)}_{\mathrm{learn}}$ ) are modulated by a function $\phi(\bar{\mathbf{p}})$ derived from the pooled representation of the text prompt.
Effect: This allows the model to focus its attention on specific object categories described in the text (e.g., "finding ships in the harbor"), effectively filtering out irrelevant background clutter and adapting to varying object densities.

2. Text-Aware Cross-Attention

The decoder utilizes a multi-head cross-attention mechanism where the modulated queries attend to both visual features (from the image encoder) and semantic features (from the text encoder).

Input: Visual tokens $\mathbf{enc}(\mathbf{I})$ and text tokens $\mathbf{emb}(\mathbf{P})$ are concatenated to form the Keys and Values for the attention mechanism.
Benefit: This enables the model to align spatial image features with the semantic intent of the prompt, improving detection accuracy in complex scenes.

3. Prediction and Decoding

Angle Encoding: To resolve angular discontinuity, the rotation angle $\theta$ is encoded using sine and cosine functions ( $\sin \theta, \cos \theta$ ) rather than direct regression.
Class Masking: A class mask is applied to the output logits during training and inference. It filters out irrelevant classes not specified in the prompt, ensuring the model only predicts objects relevant to the user's query.
Loss Function: The training objective combines classification loss, bounding box regression (L1, GIoU for OBBs, and smooth angle loss), and a Contrastive Denoising (CDN) auxiliary loss to help the model recover from noisy positive queries. Hungarian matching is used to align predictions with ground truth.

Key Contributions

Language-Guided Framework: LOGOS is the first approach to integrate textual prompts directly into the oriented object detection pipeline for remote sensing, allowing dynamic focus on specific object categories.
Prompt-Modulated Queries: The introduction of FiLM-based modulation for content queries addresses the inefficiency of fixed-query models in scenes with varying object densities.
Resolution of Ambiguity: By utilizing sine/cosine angle encoding and prompt-guided class masking, the method mitigates angular periodicity issues and reduces prediction redundancy.
State-of-the-Art Performance: The approach demonstrates superior performance across multiple benchmarks, particularly in densely packed and rotated object scenarios.

Experimental Results

The authors evaluated LOGOS on three versions of the DOTA dataset (v1.0, v1.5, and v2.0), comparing it against numerous state-of-the-art methods (e.g., Oriented R-CNN, R3Det, Deformable DETR-O, AO2-DETR).

DOTA-v1.0: LOGOS achieved a mAP of 81.32%, outperforming the second-best method (Oriented R-CNN at 80.87%). It showed significant gains in categories like Plane, Storage Tank, Harbor, and Roundabout.
DOTA-v1.5: The model achieved a mAP of 69.97%, with notable improvements in detecting Traffic Cones (94.60%) and Harbors (91.98%).
DOTA-v2.0: LOGOS reached a mAP of 66.04%, maintaining top-tier performance across diverse categories including Planes, Buildings, and Large Vehicles.

Qualitative Analysis: Visual results indicate that LOGOS effectively handles dense scenes (e.g., harbors with multiple ships) and diverse scales. However, the authors acknowledge limitations in extremely crowded scenes with small, overlapping vehicles and in detecting objects at extreme or unusual orientations where confidence drops.

Significance and Claims

The paper claims that LOGOS represents a significant step forward in the robustness and scalability of oriented object detection for remote sensing applications. By leveraging textual prompts, the method offers a flexible solution to the "fixed query" bottleneck and the "angular ambiguity" problem inherent in previous models.

The authors position this work as a bridge between Vision-Language Models and geospatial analysis, suggesting that guiding detection with natural language can significantly improve accuracy in complex, cluttered environments typical of satellite and drone imagery. They note that while the method is promising, future work is needed to handle extreme orientations and integrate multimodal data (e.g., multispectral imagery) for further robustness.