acceptodds
Under review as a conference paper at ICLR 2027

A New Dataset and Research Approach for Movable Target Detection in Airports

Abstract

Movable target detection on airport surfaces is fundamental to intelligent airport surveillance, yet the field lacks efficient detection algorithms and dedicated benchmark datasets. To address this gap, this paper introduces AFTD, a foreground target detection dataset for three types of movable airport targets: airplanes, vehicles, and persons. Through autonomous data acquisition and web data collection, over 200000 images were obtained, from which 10050 images were selected based on diversity to form the AFTD dataset. It contains 26968 airplane instances, 24759 vehicle instances, and 5064 person instances. AFTD covers diverse real-world challenges and provides fine-grained annotations, including three-level scale and occlusion annotations, as well as eight-level viewpoint annotations. Based on AFTD, we conducted extensive comparative experiments and confirmed that foundational detection algorithms are not robust to real-world scenarios. Subsequently, we propose a scene-specific prior-based approach for airport detection, along with its preliminary implementation. Furthermore, we demonstrate the cross-scenario generalizability of this research approach. AFTD can be downloaded from http://www.agvs-caac.com/aftd/aftd.html, and NSAD from https://github.com/AFTD-NSAD/AFTD-NSAD.

Primary area:
datasets and benchmarks
Keywords:
Airport Traffic, Foreground Target, Object Detection, Intelligent Airport

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_HuJmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces AFTD (Airport Foreground Target Detection), a dataset of 10,050 images for movable target detection in airports, covering airplanes, vehicles, and persons. The images come from cameras at five airports and online sources. The dataset is designed to support research and benchmarking under super multiscale, occlusion change, and viewangle change conditions. It also provides more detailed annotations for airplanes, including scale level, occlusion degree, and approximate viewangle. The authors evaluate 22 detectors and interpret their lower AP on AFTD compared with COCO val2017 as evidence that foundational algorithms lack robustness in real-world scenes. Then they proposes a scene-specific prior based approach, implemented through NSAD with nonparametric scale modeling, dynamic anchor initialization, and scale-normalized regression. The authors report performance gains over D-FINE on both AFTD and COCO val2017, with scale priors refitted for COCO. Based on these results, they conclude that the approach is effective and has cross-scenario generalizability.

Critical strengths
  1. The paper collects real-world data from an application where data access is difficult and provides detailed annotations that require substantial effort. The authors collected data at five airports using fixed and PTZ cameras over about twenty months. They also included images from online sources to cover changes in scale, occlusion, viewangle, weather, and illumination. The airplane attribute labels go beyond standard bounding boxes. In particular, occlusion degree is estimated by aligning a 3D airplane model with each instance. These annotations help distinguish AFTD from other datasets, add practical value, and reflect the substantial annotation effort. Given the barriers to data collection at airports, the data and its detailed annotations are the paper’s most important contribution.

  2. The paper evaluates a broad range of baselines and provides public links to the data and code. The authors evaluate 22 detectors on AFTD and public datasets, covering YOLO, DETR, two-stage, and anchor-free methods. These results help show how different detection architectures perform in airport scenes. Together with the publicly available data and code, they provide a useful reference for future research on airport surveillance.

Critical weaknesses
  1. The benchmark protocol is incomplete, and the key attribute annotations and challenge subsets lack quantitative evaluation. The paper does not clearly state the sizes of the AFTD training, validation, and test sets or how they are split. It is unclear whether any airports are held out for testing or how near-duplicate images from the same camera or nearby time points are handled. AFTD provides airplane annotations for scale level, occlusion degree, and approximate viewangle. It also highlights six challenges: super multiscale, occlusion change, viewangle change, illumination change, weather change, and complex background. However, the paper lacks quantitative results grouped by these attributes and challenges. The criteria and sizes of subsets for weather, illumination, and other conditions are also unclear. The authors mainly interpret the lower AP on AFTD compared with COCO val2017 as evidence that methods lack robustness for real-world scenarios, without further examining performance or differences between methods under each challenge. Although the dataset is the paper’s main contribution, the current experiments mainly show its overall difficulty. They do not fully demonstrate the value of its detailed annotations and challenge design for robustness research.

  2. NSAD lacks quantitative results from ablations and a clear baseline comparison. The paper does not report complete quantitative ablation results for dynamic anchor initialization and scale-normalized regression. It also does not clearly identify the detector and model version used as the basis for NSAD or explain how its configuration differs from D-FINE. This makes it difficult to assess the actual gains from each design. The gains on COCO after refitting the scale distribution of all classes support the approach’s use across datasets. However, the claim of cross-scenario generalizability needs to be stated more clearly. The experiments show that the approach remains effective after its priors are re-estimated on a new dataset. They do not establish that a trained model can generalize to unseen scenes.

Questions for authors
  1. A more complete description of the data split would help readers understand the benchmark results. Please report the sizes of the AFTD training, validation, and test sets and explain how they are split. This should include whether any airports are held out for testing and how near-duplicate images from the same camera or nearby time points are handled.
  2. The detailed annotations are an important contribution of the dataset. Reporting quantitative results by scale level, occlusion degree, and approximate viewangle would therefore be valuable. Please also clarify the definitions, overlap, and sample counts of the six subsets, along with whether their labels are publicly available. This would help demonstrate the dataset’s value for robustness research and make the link between the conclusions and experimental results clearer.
  3. The appendix already describes the hyperparameter search and selection process in detail. It would help to identify the detector and model version used as the basis for NSAD and explain how its configuration relates to D-FINE. Reporting separate quantitative ablation results for dynamic anchor initialization and scale-normalized regression would also help clarify the gains from each design.
  4. The meaning of approximate viewangle needs further clarification. Under the current definition, it describes the orientation of the airplane’s axis in the image plane. At low depression angles, projection effects may cause airplanes viewed from an oblique angle to be labeled as Left or Right. Discussing this limitation and distinguishing the 2D orientation distribution in Figure 9b from 3D viewing-angle coverage would make the meaning of the annotations clearer.
  5. AI use statement was not found in the provided main text or appendix. Please check the submission requirements and add a statement on whether generative AI tools were used and for what purposes, as required. Please also make the in-text citations and reference list consistent with the conference format.
Reviewer_8kPSOpenReview ↗
Rating
3
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces a new dataset for movable target detection in airport environments, named AFTD, along with a scene-prior-based object detection method called NSAD. AFTD contains approximately 20,000 images of airport scenes, covering three major object categories: airplanes, vehicles, and persons. It also provides fine-grained annotations, including airplane viewpoints, occlusion levels, and object scales. The authors evaluate 22 representative object detection algorithms on AFTD to investigate their performance and robustness in airport environments. Furthermore, NSAD is developed based on the D-FINE detector by modeling category-dependent object scale distributions and incorporating scale priors into detector optimization. Experiments on AFTD and COCO demonstrate certain improvements in detection performance.

Overall, this paper addresses a practically relevant problem and introduces a dataset with considerable scale and scene diversity. However, the methodological novelty of NSAD is limited, and the experimental validation remains insufficient, particularly regarding comparisons with existing scale-prior-based detection methods and evaluations of cross-domain generalization.

Critical strengths
  1. Practical relevance. This paper addresses movable target detection in airport environments, where significant scale variations, occlusions, and complex environmental conditions pose challenges to existing object detectors. The problem has practical significance and potential real-world applications.

  2. A new airport object detection dataset. AFTD contains approximately 20,000 images with fine-grained annotations, including object scales, viewpoints, and occlusion levels. This dataset could provide a useful benchmark for studying object detection under challenging airport scenarios.

  3. Extensive benchmark evaluation. The authors evaluate 22 representative object detection algorithms covering different detection architectures and compare their performance on AFTD and COCO. These experiments help reveal the limitations of existing detectors in airport-specific scenarios.

Critical weaknesses

1. Limited methodological novelty of NSAD.

The use of object scale priors has been widely explored in object detection. For example, YOLOv3/v4 employ K-means clustering for anchor initialization, while GADet (IEEE Sensors Journal, 2024) introduces a category-related Physical Diagonal Length Constraint (PDLC). The authors do not sufficiently clarify how NSAD fundamentally differs from these existing approaches.

2. Insufficient validation of generalization ability.

NSAD relies on dataset-specific scale distributions, which may vary across airports, camera viewpoints, and shooting distances. Evaluating models separately trained on AFTD and COCO does not adequately demonstrate cross-domain generalization.

3. Insufficient ablation studies.

The paper lacks systematic ablations on scale prior modeling strategies and hyperparameter sensitivity. Comparisons with simpler scale-prior-based approaches are also missing, making it difficult to identify the actual source of performance improvements.

4. Limited validation of fine-grained annotations.

Although AFTD provides additional annotations for viewpoints, occlusion levels, and object scales, their practical value is insufficiently demonstrated. The paper lacks detailed evaluations under these conditions and does not clearly establish how these annotations benefit research beyond conventional object detection.

Questions for authors
  1. Methodological novelty: Could the authors clarify the fundamental differences between NSAD and conventional scale-prior-based approaches, such as K-means-based anchor initialization in YOLOv3 and YOLOv4? Furthermore, how does NSAD differ from the category-related geometric prior employed by GADet (IEEE Sensors Journal, 2024), particularly its Physical Diagonal Length Constraint (PDLC)? What are the specific technical contributions that distinguish NSAD from these existing methods?

  2. Scale prior construction and generalization: In Table 3, are the category-dependent scale priors for AFTD and COCO estimated separately from their respective training sets? Does NSAD require re-estimating the scale priors for each target dataset during evaluation? If so, how do the current experiments support the claimed cross-domain generalization ability?

  3. Effectiveness of the proposed method: Could the authors clarify whether the performance improvements of NSAD mainly arise from incorporating scale priors in general or from the specific prior modeling and optimization mechanisms proposed in this paper? Do the existing experimental results provide evidence that distinguishes these two contributions?

  4. Value of fine-grained annotations: Could the authors further clarify the unique research value of the additional annotations in AFTD compared with conventional bounding-box annotations? Do the current experiments provide evidence that these annotations offer meaningful benefits for airport object detection research?

Reviewer_YJ9YOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces a new dataset, AFTD (Airport Foreground Target Detection), for detecting airplanes, vehicles, and persons, which are commonly observed at airports. It contains 10,050 images with three classes (26,968 airplanes, 24,759 vehicles, and 5,064 persons). The main characteristics of AFTD are its annotations for three object scales, occlusion levels, and eight levels of object orientation. Through experiments, the authors show that even well-known object detectors, including models in the DETR family such as RF-DETR, often fail to detect objects in this dataset, especially small ones. In addition, the paper proposes NSAD (Nonparametric Scale-Based Airport Detection), a scene-specific prior-based approach for airport target detection that uses co-occurrence and shape priors. The authors show that using this prior information can improve overall detection performance.

Critical strengths

The proposed AFTD dataset appears to contain a large number of instances for each target class, especially airplanes (26,968 instances), and to have good-quality annotations.

Critical weaknesses
  1. In general, this reviewer welcomes new datasets for the AI community. However, when evaluating a dataset contribution, it is important to consider whether the dataset offers new insights or contributions that have not been adequately covered by previous benchmarks—for example, by enabling research on a new problem or providing data that was previously insufficient for training and evaluation. In this regard, this reviewer is still unsure about the core contribution of this paper. To demonstrate the value of the dataset, the authors should have conducted at least cross-dataset evaluations, such as training on MS COCO and testing on AFTD, and vice versa. Without such results, it is difficult to assess the value of this dataset. Is this dataset really needed to address this problem?

  2. Another core contribution of this paper is the use of prior information to improve object detector performance. The authors also claim that the approach is general, since using this prior information improves detection performance not only on the proposed airport dataset (AFTD) but also on COCO-val 2017. However, to support this claim of generality, the authors should have evaluated the approach with different object detector families, such as YOLO, R-CNN, and DETR. The proposed NSAD algorithm appears to be applicable as a plug-in to different detectors. This would be a significant advantage, but the paper does not demonstrate it.

Questions for authors

In Table 3, which detector is used to obtain the NSAD results? It appears that NSAD is applied to D-FINE. Is that correct?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.