Object Detection In Aerial Images Research Articles

Classification of objects in images with distortions based on a two-stage topological analysis

The authors propose a method for automatic classification of spatial objects in images under conditions of a limited data set. The stability of the method to distortions appearing in images due to natural phenomena and partial overlap of urban infrastructure objects is investigated. High classification accuracy, when using existing approaches, requires a large training sample, including data sets with distortions, which significantly increases computational complexity. The paper proposes a method for a two-step topological analysis of images. Topological features are initially extracted by analyzing the image in the brightness range from 0 to 255, and then from 255 to 0. These features complement each other and reflect the topological structure of the object. Under certain deformations and distortions, the object preserves its structure in the form of extracted features. The advantage of the method is a small number of patterns, which reduces the computational complexity of training compared to neural networks. The proposed method is investigated and compared with the modern neural network approach. The study was performed on a DOTA dataset (Dataset for Object deTection in Aerial images) containing images of spatial objects of several classes. In the absence of distortion in the image, the neural network approach showed a classification accuracy of over 98 %, while the proposed method achieved about 82 %. Further distortions such as 90 degree rotation, 50 % narrowing and 50 % edge truncation and their combinations were applied in the experiment. The proposed method showed its robustness and outperformed the neural network approach. In the most difficult combination of the test, the decrease in classification accuracy of the neural network was 46 %, while the described method showed 12 %. The proposed method can be applied in cases with a high probability of distortion in the images. Such distortions arise in the field of geoinformatics when analyzing objects of various scales, under different weather conditions, partial overlap of one object with another, in the presence of shadows, etc. It is possible to use the proposed method in vision systems of industrial enterprises for automatic classification of the parts that belong to superimposed objects.

MSODANet: A network for multi-scale object detection in aerial images using hierarchical dilated convolutions

The object detection in aerial images is one of the most commonly used tasks in the wide-range of computer vision applications. However, the object detection is more challenging due to the following issues: (a) the pixel occupancy vary among the different scales of objects, (b) the distribution of objects is not uniform in aerial images, (c) the appearance of an object varies with different view-points and illumination conditions, and (d) the number of objects, even though they belong to same type, vary across the images. To address these issues, we propose a novel network for multi-scale object detection in aerial images using hierarchical dilated convolutions, called as mSODANet. In particular, we probe hierarchical dilated network using parallel dilated convolutions to learn the contextual information of different types of objects at multiple scales and multiple field-of-views. The introduced hierarchical dilated network captures the visual information of aerial image more effectively and enhances the detection capability of the model. Further, the extensive experiments conducted on three challenging publicly available datasets, i.e., Visdrone2019, DOTA (OBB & HBB), NWPU VHR-10, demonstrate the effectiveness of the proposed mSODANet and achieve the state-of-the-art performance on all three datasets.

Enhancing object detection in aerial images.

Unmanned Aerial Vehicles have proven to be helpful in domains like defence and agriculture and will play a vital role in implementing smart cities in the upcoming years. Object detection is an essential feature in any such application. This work addresses the challenges of object detection in aerial images like improving the accuracy of small and dense object detection, handling the class-imbalance problem, and using contextual information to boost the performance. We have used a density map-based approach on the drone dataset VisDrone-2019 accompanied with increased receptive field architecture such that it can detect small objects properly. Further, to address the class imbalance problem, we have picked out the images with classes occurring fewer times and augmented them back into the dataset with rotations. Subsequently, we have used RetinaNet with adjusted anchor parameters instead of other conventional detectors to detect aerial imagery objects accurately and efficiently. The performance of the proposed three step pipeline of implementing object detection in aerial images is a significant improvement over the existing methods. Future work may include improvement in the computations of the proposed method, and minimising the effect of perspective distortions and occlusions.

Rethinking Classification of Oriented Object Detection in Aerial Images

With the help of the rapid development of technology, especially the prevalence of UAVs (unmanned aerial vehicles), object detection in aerial images gains much more attention in computer vision and deep learning. However, traditional methods use horizontal bounding boxes for object representation leading to inconsistency between objects and features. Therefore, many detectors are being built to tackle this problem, and normally they use the conventional approaches of training and testing to achieve the results. Our pipeline proposed to strengthen not only the classification but also localization via independent training processes using convex-hull transformation in data pre-processing phase. We experimented with the well-designed S2ANet, R3Det, ReDet, RoI Transformer and Oriented R-CNN on the well-established oriented object detection dataset DOTA. Then we adopt the best detectors with the well-known classification network EfficientNet to our proposed pipeline and achieve promising results on the oriented object detection DOTA dataset. Moreover, our pipeline can flexibly be adapted to various oriented object detection baselines improving the results in classification via independent extensive training cycles.

Gaussian Focal Loss: Learning Distribution Polarized Angle Prediction for Rotated Object Detection in Aerial Images

With the increasing availability of aerial data, object detection in aerial images has aroused more and more attention in remote sensing community. The difficulty lies in accurately predicting the angular information for each target when using the oriented bounding boxes to represent the arbitrary oriented objects, as the periodicity of the angle could cause inconsistency between target angle values. To resolve the problem, recent works propose to perform angular prediction from a regression problem to a classification task with circular smooth label. However, we find that current loss functions applying to binary soft labels need to approximate the soft label values at each position. When summed over all the negative angle categories, these relatively insignificant loss values can overwhelm the target angle category, thus preventing the network from predicting precise angle information. In this paper, we propose a novel loss function that acts as a more effective alternative to the classification-based rotated detectors. By constructing the classification loss with adaptive Gaussian attenuation on the negative locations, our training objective can not only avoid discontinuous angle boundaries but also enable the network to obtain more accurate angle predictions with higher response at peaks. Moreover, an aspect ratio-aware factor was proposed based on our loss function to enhance the robustness of the model for determining the orientation for square-like objects. Extensive experiments on aerial image datasets DOTA, HRSC2016, and UCAS-AOD demonstrated the effectiveness and superior performances of our approaches.

DARDet: A Dense Anchor-Free Rotated Object Detector in Aerial Images

Rotated object detection in aerial images has received increasing attention for a wide range of applications. However, it is also a challenging task due to the huge variations of scale, rotation, aspect ratio, and densely arranged targets. Most existing methods heavily rely on a large number of pre-defined anchors with different scales, angles, and aspect ratios, and are optimized with a distance loss. Therefore, these methods are sensitive to anchor hyper-parameters and easily suffer from performance degradation caused by boundary discontinuity. To handle this problem, in this paper, we propose a dense anchor-free rotated object detector (DARDet) for rotated object detection in aerial images. Our DARDet directly predicts five parameters of rotated boxes at each foreground pixel of feature maps. We design a new alignment convolution module to extracts aligned features and introduce a PIoU loss for precise and stable regression. Our method achieves state-of-the-art performance on three commonly used aerial objects datasets (i.e., DOTA, HRSC2016, and UCAS-AOD) while keeping high efficiency. Code is available at https://github.com/zf020114/DARDet.

R2YOLOX: A Lightweight Refined Anchor-Free Rotated Detector for Object Detection in Aerial Images

Existing anchor-based rotated object detection methods have achieved some amazing results, but these methods require some manual preset anchors, which not only introduce additional hyperparameters but also introduce extra computational burdens. Due to the above drawbacks, anchor-free methods have been rapidly developed in recent years. However, the existing high-performance anchor-free rotated object detection methods are relatively complex and the inference speed is also slow. And Yolo series models not only maintain high-efficiency inference but also keep competitive performance detection performance in the general object detection tasks. Hence, we propose an anchor-free rotated detector based on the YOLOX method for object detection in aerial images. Our methods consist of two improvements: a Refined Rotated Module (RRM) and a new assigner method which is called the Gaussian distribution Sampling Optimal Transport Assignment method (GSOTA). The RRM can align features and get more useful priors for final detector heads. The GSOTA uses Gaussian Distribution to model the oriented bounding box (OBB) firstly, and a Gaussian Center Sampling method (GCS) with maximum classification center mean (MCCM) is proposed to simplify the label Assignment Optimal Transport problem, finally using an improved dynamic top-k strategy to get an approximate solution. Extensive experiments demonstrate that our models can achieve competitive performance in several challenging aerial object detection datasets while keeping the best efficiency. Our R2YOLOX-X model achieves 79.33%, 97.4%, 97.7%, and 92.5% mAP on the DOTA, HRSC2016, UCAS-AOD, and FGSD2021, respectively, while R2YOLOX-S can reach the fastest 58.2 FPS when inferencing on aerial datasets and R2YOLOX-L gets the best speed-accuracy trade-off.

Multi-Oriented Object Detection in Aerial Images With Double Horizontal Rectangles.

Most existing methods adopt the quadrilateral or rotated rectangle representation to detect multi-oriented objects. Yet, the same oriented object may correspond to several different representations, due to different vertex ordering, or angular periodicity and edge exchangeability. To ensure the uniqueness of the representation, some engineered rules are usually added. This makes these methods suffer from discontinuity problem, resulting in degraded performance for objects around some orientation. In this article, we propose to encode the multi-oriented object with double horizontal rectangles (DHRec) to solve the discontinuity problem. Specifically, for an oriented object, we arrange the horizontal and vertical coordinates of its four vertices in left-right and top-down order, respectively. The first (resp. second) horizontal box is given by two diagonal points with smallest (resp. second) and third (resp. largest) coordinates in both horizontal and vertical dimensions. We then regress three factors given by area ratios between different regions, helping to guide the oriented object decoding from the predicted DHRec. Inherited from the uniqueness of horizontal rectangle representation, the proposed method is free of discontinuity issue, and can accurately detect objects of arbitrary orientation. Extensive experimental results show that the proposed method significantly improves the existing baseline representation, and outperforms state-of-the-art methods. The code is available at: https://github.com/lightbillow/DHRec.

An Anchor-Free Network With Density Map and Attention Mechanism for Multiscale Object Detection in Aerial Images

Accurate detection of the multiple classes in aerial images has become possible with the use of anchor-based object detectors. However, anchor-based object detectors place a large number of preset anchors on images and regress the target bounding box while anchor-free object detections predict the location of objects directly and avoid the carefully predefined anchor box parameters. Object detection in aerial images is faced with two main challenges: 1) the scale diversity of the geospatial objects; and 2) the cluttered background in complex scenes. In this letter, to address these challenges, we present a novel Anchor-Free Network with a Density map and attention mechanism (DA <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> FNet). Considering the extreme density variations of the detection instances among the different categories in aerial images, the proposed DA <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> FNet model conducts density map estimation with image-level supervision for the geospatial object counting, to acquire global knowledge about the scale information. A simple and effective image-level global counting loss function is also introduced. In addition, a compositional attention network is further introduced to enhance the saliency of the foreground objects. The proposed DA <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> FNet method was compared with the state-of-the-art object detection models, achieving excellent performance on the NWPU VHR-10, RSOD, and DOTA datasets.

MRDet: A Multihead Network for Accurate Rotated Object Detection in Aerial Images

Objects in aerial images usually have arbitrary orientations and are densely located over the ground, making them extremely challenge to be detected. Many of the recent developed methods attempt to solve these issues by estimating an extra orientation parameter and placing dense anchors, which will result in high model complexity and computational costs. In this article, we propose an arbitrary-oriented region proposal network (AO-RPN) to generate oriented proposals transformed from horizontal anchors. The AO-RPN is very efficient with only a few amounts of parameters increase than the original RPN. Furthermore, to obtain accurate bounding boxes, we decouple the detection task into multiple subtasks and propose a multihead network to accomplish them. Each head is specially designed to learn the features optimal for the corresponding task, which allows our network to detect objects accurately. We name it multihead rotated object detector (MRDet). We evaluate the performance of the proposed MRDet on two challenging benchmarks, i.e., DOTA and HRSC2016, and compare it with several state-of-the-art methods. Our method achieves very promising results, which clearly demonstrates its effectiveness. Code has been available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/qinr/MRDet</uri> .

Probability Differential-Based Class Label Noise Purification for Object Detection in Aerial Images

Modern object detection for aerial images requires numerous annotated data. However, the data annotation process inevitably introduces noise due to the bird’s eye view perspective of aerial images and the professional requirements of annotations. While recent noise-robust object detection methods achieved great success, the noise side-effect during the early training stage was still a problem. As demonstrated in this letter, noise during the early training stage will cumulatively affect the final performance. Based on the above observations, we propose a training strategy called correction maximization training to purify the noisy annotations and then train models. Specifically, we design a novel noise filter called the probability differential (PD) to identify and revise wrong labels. After purification, we train the detector with the revised dataset. Compared to existing works, the proposed method could be adapted in most modern object detectors (e.g., Faster RCNN and RetinaNet) and requires little hyperparameter tuning across different datasets and models. Extensive experiments on DOTA show that the proposed method achieves state-of-the-art results with both symmetric and asymmetric noise.

Learning Calibrated-Guidance for Object Detection in Aerial Images

Object detection is one of the most fundamental yet challenging research topics in the domain of computer vision. Recently, the study on this topic in aerial images has made tremendous progress. However, complex background and worse imaging quality are obvious problems in aerial object detection. Most state-of-the-art approaches tend to develop elaborate attention mechanisms for the space-time feature calibrations with arduous computational complexity, while surprisingly ignoring the importance of feature calibrations in channel-wise. In this work, we propose a simple yet effective calibrated-guidance (CG) scheme to enhance channel communications in a feature transformer fashion, which can adaptively determine the calibration weights for each channel based on the global feature affinity correlations. Specifically, for a given set of feature maps, CG first computes the feature similarity between each channel and the remaining channels as the intermediary calibration guidance. Then, rerepresenting each channel by aggregating all the channels weighted together via the guidance operation. Our CG is a general module that can be plugged into any deep neural networks, which is named as CG-Net. To demonstrate its effectiveness and efficiency, extensive experiments are carried out on both oriented object detection task and horizontal object detection task in aerial images. Experimental results on two challenging benchmarks (<i>i</i>.<i>e</i>., DOTA and HRSC2016) demonstrate that our CG-Net can achieve the new state-of-the-art performance in accuracy with a fair computational overhead. The source code has been open sourced at <uri>https://github.com/WeiZongqi/CG-Net</uri>.

FADA: Feature Aligned Domain Adaptive Object Detection in Remote Sensing Imagery

Deep learning-based object detectors have been widely adopted in the field of remote sensing imagery interpretation. These detectors heavily depend on the expensive large-scale labeled datasets, while the scarce remote sensing datasets limit the performance. The domain adaptive object detection can alleviate this problem. However, it struggles with the confusing feature’s alignment, damaging the domain generalization performance, especially for the remote sensing scene with sparse objects and diverse backgrounds. For that reason, a semisynthetic data generator (SDG) is proposed to automatically generate the remote sensing dataset with low cost and replace the real-world training dataset, a <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">feature aligned domain adaptive object detector</i> (FADA) is proposed to enhance the domain adaptation among the cross-domain remote sensing images. The FADA contains two proposed modules in addition to the base detector: an adversarial-based foreground alignment (AFA) and a prototype-based confusing feature alignment (PCFA). The AFA aligns the cross-domain foreground feature by adversarial training (AT), and it can filter the noisy background feature that is not suitable to transfer. Then, the PCFA adaptively aligns the confusing background and foreground feature, further promoting the domain adaptation performance. Comprehensive experiments validate the effectiveness of the proposed method. Compared with the baseline model trained on the semisynthetic source dataset, our FADA improves the generalized performance on the real-world target dataset a large-scale Dataset for Object deTection in Aerial images (DOTA) by 15.7% average precision (AP) and achieves state-of-the-art results.

Reducing False Detections in Aerial Images by Exploiting the Context Information and Centroid Relationship

Object detection in aerial images is critical for military and civilian applications. Due to its broad imaging range, an aerial image contains more complex scenes than natural images, which may result in increased confusion between targets and the background. To address this issue, we propose a context-aware and centroid-prior-based object detector for the object detection task in aerial images. Unlike existing object detectors, the proposed model investigates the feasibility of reducing false detections from confusing backgrounds by exploiting both semantic and neighbor contexts. Additionally, a new label assignment algorithm is developed based on the centroid prior assumption to mitigate the localization errors introduced by context information. The experimental results demonstrate that the proposed model outperforms six advanced object detectors on the optical and synthetic aperture radar (SAR) image datasets, indicating the proposed model&#x2019;s effectiveness.

Multiscale Semantic Fusion-Guided Fractal Convolutional Object Detection Network for Optical Remote Sensing Imagery

Optical remote sensing object detection is a challenging task, because of the complex background interference, ambiguous appearances of tiny objects, densely arranged circumstances, and multiclass object with vaster scale variances and irregular aspect ratios. The performance of object detection is seriously restricted. Thus, in this article, inspired by the anchor-free object detection framework, and aiming to solve these difficulties to improve the optical remote sensing object detection performance, a powerful one-stage detector of multiscale semantic fusion-guided fractal convolution network (MSFC-Net) is proposed. First, facing these strong-coupled semantic relations in each complex scene, a compound semantic feature fusion (CSFF) way is designed for generating an effective semantic description, which is a benefit to pixel-wise object center point interpretation. In addition, it can be easily extended into a semantic segmentation task. Second, in view of accurate multiclass pixel-wise center point predictions based on an effective compound semantic description, a novel fractal convolution (FC) regression layer is designed, which adaptively achieves the regression of multiscale bounding boxes (bboxes) with irregular aspect ratio under no priori information. Third, related to the set up FC regression layer, a specific hybrid loss is designed to make the proposed MSFC-Net converge better. Finally, the extensive experiments on challenge data sets of large-scale dataset for object detection in aerial images (DOTA) and object detection in optical remote sensing images (DIOR) datasets are carried out, and comparisons indicate that the proposed MSFC-Net can perform the remarkable performance than other state-of-the-art one-stage detectors, as it can reach 80.26&#x0025; mean average precision (mAP) and 79.33&#x0025; mF1 on DOTA and 70.08&#x0025; mAP and 73.45&#x0025; mF1 on DIOR. Then, our work is available at <uri>https://github.com/ZhAnGToNG1/MSFC-Net</uri>.

Rotated Object Detection via Scale-Invariant Mahalanobis Distance in Aerial Images

Rotated object detection in aerial images is a meaningful yet challenging task as objects are densely arranged and have arbitrary orientations. The eight-parameter (coordinates of box vectors) methods in rotated object detection usually use ln-norm losses (L1 loss, L2 loss, and smooth L1 loss) as loss functions. As ln-norm losses are mainly based on non-scale-invariant Minkowski distance, using ln-norm losses will lead to inconsistency with the detection metric rotational Intersection-over-Union (IoU) and training instability. To address the problems, we use Mahalanobis distance to calculate loss between the predicted and the target box vertices' vectors, proposing a new loss function called Mahalanobis Distance Loss (MDL) for eight-parameter rotated object detection. As Mahalanobis distance is scale-invariant, MDL is more consistent with detection metric and more stable during training than ln-norm losses. To alleviate the problem of boundary discontinuity like all other eight-parameter methods, we further take the minimum loss value to make MDL continuous at boundary cases. We achieve state-of-art performance on DOTA-v1.0 with the proposed method MDL. Furthermore, compared to the experiment that uses smooth L1 loss, we find that MDL performs better in rotated object detection.

Object Detection for Aerial Images With Feature Enhancement and Soft Label Assignment

Object detection in aerial images, different from general object detection, faces with several challenges such as arbitrary-oriented objects and extremely imbalanced foreground-background distribution. Although some recent proposed aerial object detection methods achieve promising results, they are mainly anchor-based detectors which rely heavily on pre-defined anchor boxes and the final detection performance is sensitive to anchor-related hyper-parameters. In contrast, in this paper, we present an anchor-free detector with feature enhancement and soft label assignment (FSDet) which adopts a simpler design and achieves competitive performance. Specifically, to address the feature misalignment for detecting oriented objects, we propose an oriented feature refinement module to align the features with oriented objects. To alleviate the background issue, we design a class-aware context aggregation module to integrate the intra-class context information and suppress background context. Moreover, we propose a soft label assignment mechanism to measure the weight of training samples within the arbitrary-oriented objects, which can concentrate more on representative items with regard to their potential to detect oriented objects, achieving a more stable optimization during training. Extensive experiments on several datasets suggest that the proposed method is superior to the state-of-the-art methods and achieves a better trade-off between speed and accuracy.

Anchor Retouching via Model Interaction for Robust Object Detection in Aerial Images

Object detection has made tremendous strides in computer vision. Small object detection with appearance degradation is a prominent challenge, especially for aerial observations. To collect sufficient positive/negative samples for heuristic training, most object detectors preset region anchors in order to calculate intersection-over-union (IoU) against the ground-truth data. In this case, small objects are frequently abandoned or mislabeled. In this article, we present an effective dynamic enhancement anchor network (DEA-Net) to construct a novel training sample generator. Different from the other state-of-the-art (SOTA) techniques, the proposed network leverages a sample discriminator to realize interactive sample screening between an anchor-based unit and an anchor-free unit to generate eligible samples. Besides, multi-task joint training with a conservative anchor-based inference scheme enhances the performance of the proposed model while reducing computational complexity. The proposed scheme supports both oriented and horizontal object detection tasks. Extensive experiments on two challenging aerial benchmarks (i.e., Dataset of Object deTection in Aerial images (DOTA) and HRSC2016) indicate that our method achieves SOTA performance in accuracy with moderate inference speed and computational overhead for training. On DOTA, our DEA-Net which integrated with the baseline of RoI-transformer surpasses the advanced method by 0.40&#x0025; mean-average-precision (mAP) for oriented object detection with a weaker backbone network (ResNet-101 vs. ResNet-152) and 3.08&#x0025; mAP for horizontal object detection with the same backbone. Besides, our DEA-Net which integrated with the baseline of ReDet achieves the SOTA performance by 80.37&#x0025;. On HRSC2016, it surpasses the previous best model by 1.1&#x0025; using only three horizontal anchors. The source code and the training set are made publicly available at <uri>https://github.com/QxGeng/DEA-Net</uri>.

Fine-Grained Feature Enhancement for Object Detection in Remote Sensing Images

Recently, object detection in aerial images has ushered in a new challenge—a new benchmark for fine-grained object recognition in high-resolution remote sensing imagery called FAIR1M has been proposed. Fine-grained categories usually have smaller inter class differences and intra-class similarities, which is more difficult to classify with existing object detectors. To address this problem, we propose two enhanced strategies on the current two-stage object detection algorithm. The first strategy uses attention-based group feature enhancement called group enhance module (GEM). By extending and grouping feature channels, the model can improve the ability to extract various discriminative features. The second strategy is to emphasize the sub-saliency feature learning, avoiding the network only focusing on the most significant part of the feature and ignoring the other parts. Our method is easy to implement and effective, and experiments show that our method can improve the Oriented regions with convolutional neural networks features (R-CNN) by about 1.45 mAP on the FAIR1M benchmark.

AVS-YOLO: Object Detection in Aerial Visual Scene

Difficult object detection and class imbalance in object detection are the two main challenges faced by aerial image object detection. Difficult objects include small objects, objects of scale variation and objects with serious background interference. Class imbalances come from the number of different classes of objects and sampling of positive and negative samples. Due to these challenges, conventional object detection models usually cannot effectively detect objects in aerial images, especially in the balance between network speed and accuracy. In this paper, the YOLOv3 network structure was improved and an object detection method under the aerial visual scene (AVS-YOLO) was proposed. By introducing a type of densely connected feature pyramid strategy, a scale-aware attention module was constructed, considering both residual dense network blocks and the median-frequency-balancing mechanism. On this basis, an algorithm with ideal speed and accuracy for object detection is obtained. To verify the effectiveness of the algorithm, AVS-YOLO and YOLOv3 were both used to test the VisDrone-DET2019 and UAVDT. The experimental results show that the AP of AVS-YOLO increases by 6.22% and 5.09% on the VisDrone2019 and UAVDT datasets, respectively, compared with YOLOv3. In addition, the AP of AVS-YOLO is 1.82% higher than that of YOLOv4 on the VisDrone2019 dataset. In terms of detection speed, AVS-YOLO can process 31.8 frames per second on a single Nvidia GTX 2080Ti GPU, compared with 44.1 frames per second for YOLOv3. Compared with the other one-stage network in the field of object detection, AVS-YOLO currently achieves the state-of-the-art performance with similar calculation amount on this dataset.