Publications

Detect Any Keypoints: An Efficient Light-Weight Few-Shot Keypoint Detector

>The 38th Annual AAAI Conference on Artificial Intelligence (AAAI 2024)
>Changsheng Lu, Piotr Koniusz
>In order to maintain flexibility of detecting varying number of keypoints, existing few-shot keypoint detection (FSKD) approaches modulate query feature map per support keypoint, then detect the corresponding keypoint from each modulated feature via a detection head. Such modulation-detection separate design would scale up model into heavy yet slow one when the number of keypoints increases. To overcome this issue, we design a novel light-weight detector which combines modulation and detection into one step, greatly reducing the computational cost without sacrifice of performance. Moreover, to bridge the large domain shift of keypoints between seen and unseen species, we further improve our model with mean feature based contrastive learning to align keypoint distributions that encourage better keypoint representations for FSKD.

Few-shot Keypoint Detection with Uncertainty Learning for Unseen Species

>IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2022)
>Changsheng Lu, Piotr Koniusz
>Current non-rigid object keypoint detectors perform well on a chosen kind of species and body parts, and require a large amount of labelled keypoints for training. Moreover, their heatmaps, tailored to specific body parts, cannot recognize novel keypoints (keypoints not labelled for training) on unseen species. In this paper, we propose a versatile Few-shot Keypoint Detection (FSKD) pipeline which could not only detect base keypoints but also the novel keypoints for unseen species.

ElDet: An Anchor-free General Ellipse Object Detector

>Asian Conference on Computer Vision (ACCV 2022, to appear)
>Tianhao Wang, Changsheng Lu, Ming Shao, Xiaohui Yuan, Siyu Xia
>Ellipse detection is a fundamental task in object shape analysis. Under complex environments, the traditional image processing based approaches may under-perform due to the hand-crated features. Instead, CNN-based approaches are more robust and powerful. In this paper, we introduce an efficient anchor-free data-augmentation based general ellipse detector, termed ElDet.

A Fast Stain Normalization Network for Cervical Papanicolaou Images

>International Conference on Neural Information Processing (ICONIP 2022, to appear)
>Jiawei Cao, Changsheng Lu, Kaijie Wu, Chaochen Gu
>The domain shift between different styles of stain images greatly challenges the generalization of computer-aided diagnosis (CAD) algorithms. To bridge the gap, color normalization is a prerequisite for most CAD algorithms. This paper designs a fast normalization network (FTNC-Net) for cervical Papanicolaou stain images based on learnable bilateral filtering.

Industrial Scene Text Detection with Refined Feature-attentive Network

>IEEE Transactions on Circuits and Systems for Video Technology (published)
>Tongkun Guan, Chaochen Gu, Changsheng Lu, Jingzheng Tu, Qi Feng, Kaijie Wu, Xinping Guan
>Detecting the marking characters of industrial metal parts remains challenging due to low visual contrast, uneven illumination, corroded surfaces, and cluttered background of metal part images. In this paper, we propose a refined feature-attentive network (RFN) to solve the inaccurate localization problem.

Arc-support Line Segments Revisited: An Efficient High-quality Ellipse Detection

>IEEE Transactions on Image Processing (T-IP) (published)
>Changsheng Lu, Siyu Xia, Ming Shao and Yun Fu
>Over the years many ellipse detection algorithms spring up and are studied broadly, while the critical issue both accurately and efficiently detecting ellipses in real-world images still remains a challenge such that we cannnot find an applicable ellipse detector from Matlab and OpenCV. The main research task of paper is to propose an accurate and efficient ellipse detecotr.

Deep Transfer Neural Network Using Hybrid Representations of Domain Discrepancy

>Neurocomputing (published)
> Changsheng Lu, Chaochen Gu, Kaijie Wu, Siyu Xia, Haotian Wang, Xinping Guan
>Transfer neural networks have been successfully applied in many domain adaptation tasks. The initiative of most of the current transfer networks, essentially, is optimizing a single distance metric between the source domain and target domain, while few studies integrate multiple metrics for training transfer networks. In this paper, we propose an architecture of transfer neural network equipped with hybrid representations of domain discrepancy, which could incorporate the advantages of different types of metrics as well as compensate their imperfections. In our architecture, the Maximum Mean Discrepancy (MMD) and $\mathcal{H}$-divergence based domain adaptations are combined for simultaneous distribution alignment and domain confusion. Through extensive experiments, we find that the proposed method is able to achieve compelling transfer performance across the datasets with domain discrepancy from small scale to large scale. Especially, the proposed method can be promisingly used to predict the viewpoint of 3D-printed workpiece even trained without labels of real images. The visualization of learned features and adapted distributions by our transfer network highlights that the proposed approach could effectively learn the similar features between two domains and deal with a wide range of transfer tasks.

PointDoN: A Shape Pattern Aggregation Module for Deep Learning on Point Cloud

>International Joint Conference on Neural Networks (IJCNN 2019) (oral presentation)
>Shuxin Zhao, Chaochen Gu, Changsheng Lu, Ye Huang, Kaijie Wu and Xinping Guan
>As point cloud is a typical and significant type of geometric 3D data, deep learning on the classification and segmentation of point cloud has received widely interests recently. However, the critical problems to process the irregularity of point cloud and feature extraction of shape pattern have not yet been fully explored. In this paper, a geometric deep learning architecture based on our PointDoN module is presented. Inspired by the Difference of Normals (DoN) in traditional point clouds processing, our PointDoN module is a feature aggregation module combining DoN shape pattern descriptor with both 3D coordinates and extra features (such as RGB colors). Our PointDoN-based architecture can be flexibly applied to multiple point cloud processing tasks such as 3D shape classification and scene semantic segmentation. Experiments demonstrate that PointDoN model achieves state-of-the-art results on multiple types of challenging benchmark datasets.

Viewpoint Estimation using Triplet Loss with A Novel Viewpoint-based Input Selection Strategy

>Journal of Physics: Conference Series 2019
>Chanjian Gu, Changsheng Lu, Chanchen Gu and Xinping Guan
>Viewpoint estimation is a fundamental procedure in vision-based robot tasks. A good viewpoint of the camera relative to objects can help the visual system perform better both in observation and manipulation. Recently, CNN-based algorithms, which can effectively extract discriminative features from images in challenging conditions are utilized to handle the viewpoint estimation problem. However, most existing algorithms focus on how to leverage the extracted deep features while neglecting the spatial relationship among images that captured from various viewpoints. In this paper, we present a deep metric learning method for solving the viewpoint estimation problem. A triplet loss with a novel viewpoint-based input selection strategy is introduced, which could learn more powerful features after incorporating the spatial relationship between viewpoints. Combined with the traditional classification loss, the presented loss can further enhance the discriminative power of features. To evaluate the performance of our method, a dataset containing a large number of images generated from five different texture-less workpieces is built and the experiment results show the effectiveness of the proposed method.

Viewpoint Estimation for Workpieces with Deep Transfer Learning from Cold to Hot

>2018 International Conference on Neural Information Processing (ICONIP) (oral presentation)
>Changsheng Lu, Haotian Wang, Chaochen Gu, Kaijie Wu, Xinping Guan
>In this paper, we propose a deep transfer learning method for viewpoint estimation by transferring priori knowledge from labeled synthetic images to unlabeled real images. The synthetic images are rendered from 3D Computer-Aided Design (CAD) models and annotated automatically. To boost the performance of deep transfer network, we design a new two-stage training strategy called cold-to-hot training. At the cold start stage, deep networks are trained for the joint tasks of classification and knowledge transfer in the absence of labels of real images. But after it turns into hot stage, the pseudo labels of real images are employed for controlling the distributions of input data.

A Spatio-Temporal Fully Convolutional Network for Breast Lesion Segmentation in DCE-MRI

>2018 International Conference on Neural Information Processing (ICONIP)
>Mingjian Chen, Hao Zheng, Changsheng Lu, Enmei Tu, Jie Yang and Nikola Kasabov
>In this paper, we propose a novel end-to-end network utilizing both spatial and temporal features for fully automated breast lesion segmentation from dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI). Our network is based on a modified convolutional neural network and a recurrent neural network, and it is capable of unearthing rich spatio-temporal features. In our network, a multi-pathway structure and a fusion operator are introduced to acquire 3D information of different tissues, which is helpful for reducing false positive segmentation while boosting accuracy.

Circle Detection by Arc-support Line Segments

>2017 IEEE International Conference on Image Processing (ICIP) (oral presentation)
>Changsheng Lu, Siyu Xia, Wanming Huang, Ming Shao and Yun Fu
>We propose a novel method for circle detection by analysing and refining arc-support line segments. The key idea is to use line segment detector to extract the arc-support line segments which are likely to make up the circle, instead of all line segments. Each couple of line segments is analyzed to form a valid pair and followed by generating initial circle set. Through the mean shift clustering, the circle candidates are generated and verified based on the geometric attributes of circle edge. Finally, twice circle fitting is applied to increase the accuracy for circle locating and radius measuring.