Label assignment (LA) is one of the essential phases in the object detection paradigm and aims to classify samples as foreground or background. Current LA strategies generally discriminate samples by explicit thresholds and then calculate weighted losses based on their significances. However, existing methods mostly neglect to consider the importance of samples comprehensively due to the uneven distribution of objects and the limitations of detector structures. In this paper, we propose a hierarchical equalization loss (HEL) by reconsidering the underlying factors affecting sample weights. First, we mitigate sample imbalance at three progressive levels. (1) Task level. We propose task-reconciled weights (TRW) to overcome the effects caused by inter-task inconsistencies (i.e., the inherent differences of classification and localization). (2) Instance level. We propose instance-aware normalization (IAN) for reconstructing the distribution of sample weights within an instance to suppress environmental noise. (3) Pyramid level. We propose hierarchical modulation (HM) to alleviate the unbalanced distribution of multi-scale objects on feature pyramids. Then, we stack the above three mechanisms and formulate the effective weighted loss. Moreover, we propose a staggered candidate bag construction (SCBC) mechanism to further improve the robustness of our method. Without adding any extra overhead, HEL can improve the performance of representative detectors by an impressive margin. Equipped with HEL, a single “ResNet-50+FPN+Head” detector can achieve a performance of 41.9 AP on COCO under 1× schedule, outperforming other existing LA methods. Extensive experiments conducted on multiple backbones and datasets demonstrate the effectiveness of our method.
26 November 2015 in Computational and Mathematical Methods in Medicine
The phacoemulsification surgery is one of the most advanced surgeries to treat cataract. However, the conventional surgeries are always with low automatic level of operation and over reliance on the ability of surgeons. Alternatively, one imaginative scene is to use video processing and pattern recognition technologies to automatically detect the cataract grade and intelligently control the release of the ultrasonic energy while operating. Unlike cataract grading in the diagnosis system with static images, complicated background, unexpected noise, and varied information are always introduced in dynamic videos of the surgery. Here we develop a VidEo-Based Intelligent Recognitionand Decision (VEBIRD) system, which breaks new ground by providing a generic framework for automatically tracking the operation process and classifying the cataract grade in microscope videos of the phacoemulsification cataract surgery. VEBIRD comprises a robust eye (iris) detector with randomized Hough transform to precisely locate the eye in the noise background, an effective probe tracker with Tracking-Learning-Detection to thereafter track the operation probe in the dynamic process, and an intelligent decider with discriminative learning to finally recognize the cataract grade in the complicated video. Experiments with a variety of real microscope videos of phacoemulsification verify VEBIRD’s effectiveness.
Recently, semantic segmentation has been widely applied in various realistic scenarios. Many semantic segmentation backbone networks use various forms of dense connection to improve the efficiency of gradient propagation in the network. They achieve excellent segmentation accuracy but lack inference speed. Therefore, we propose a backbone network SCDNet with a dual path structure and higher speed and accuracy. Firstly, we propose a split connection structure, which is a streamlined lightweight backbone with a parallel structure to increase inference speed. Secondly, we introduce a flexible dilated convolution using different dilation rates so that the network can have richer receptive fields to perceive objects. Then, we propose a three-level hierarchical module to effectively balance the feature maps with multiple resolutions. Finally, a refined flexible and lightweight decoder is utilized. Our work achieves a trade-off of accuracy and speed on the Cityscapes and Camvid datasets. Specifically, we obtain a 36% improvement in FPS and a 0.7% improvement in mIoU on the Cityscapes test set.
30 April 2024 in Multimedia Tools and Applications
Instance segmentation is a computer vision task that aims to give each pixel in an image an instance-specific label. Recently, researchers have shown growing interest in real-time instance segmentation. In this paper, we propose a novel center-based real-time instance segmentation method (CenterInst), which follows the FastInst meta-architecture. Key design aspects include a center-guided query selector, a center-guided sampling-based query decoder, and a lightweight dual-path decoder. The center-guided query selector selects queries via the per-pixel prediction of center point probabilities, avoiding excessive query proposals for single instances. The center-guided sampling-based query decoder adaptively generates local sampling points based on center positions, employing adaptive mixing to update queries without irrelevant sampling disturbances. The lightweight dual-path decoder enhances inference speed and maintains accuracy via pixel decoding on every layer during training but only utilizing the final layer’s decoder during inference. The experimental results show CenterInst achieves superior accuracy and speed compared to state-of-the-art real-time instance segmentation methods.
9 December 2019 in Lecture Notes in Computer Science