Transformer-based models have improved visual tracking, but most still cannot run in real time on resource-limited devices, especially for unmanned aerial vehicle (UAV) tracking. To achieve a better balance between performance and efficiency, we propose AVTrack, an adaptive computation tracking framework that adaptively activates transformer blocks through an Activation Module (AM), which dynamically optimizes the ViT architecture by selectively engaging relevant components. To address extreme viewpoint variations, we propose to learn view-invariant representations via mutual information (MI) maximization. In addition, we propose AVTrack-MD, an enhanced tracker incorporating a novel MI maximization-based multi-teacher knowledge distillation framework. Leveraging multiple off-the-shelf AVTrack models as teachers, we maximize the MI between their aggregated softened features and the corresponding softened feature of the student model, improving the generalization and performance of the student, especially under noisy conditions. Extensive experiments show that AVTrack-MD achieves performance comparable to AVTrack’s performance while reducing model complexity and boosting average tracking speed by over 17%. Codes is available at https://github.com/wuyou3474/AVTrack.

Various substances that possess liquid states include drinking water, various types of fuel, pharmaceuticals, and chemicals, which are indispensable in our daily lives. There are numerous real-world applications for liquid content detection in transparent containers, for example, service robots, pouring robots, security checks, industrial observation systems, etc. However, the majority of the existing methods either concentrate on transparent container detection or liquid height estimation; the former provides very limited information for more advanced computer vision tasks, whereas the latter is too demanding to generalize to open-world applications. In this paper, we propose a dataset for detecting liquid content in transparent containers (LCDTC), which presents an innovative task involving transparent container detection and liquid content estimation. The primary objective of this task is to obtain more information beyond the location of the container by additionally providing certain liquid content information which is easy to achieve with computer vision methods in various open-world applications. This task has potential applications in service robots, waste classification, security checks, and so on. The presented LCDTC dataset comprises 5916 images that have been extensively annotated through axis-aligned bounding boxes. We develop two baseline detectors, termed LCD-YOLOF and LCD-YOLOX, for the proposed dataset, based on two identity-preserved human posture detectors, i.e., IPH-YOLOF and IPH-YOLOX. By releasing LCDTC, we intend to stimulate more future works into the detection of liquid content in transparent containers and bring more focus to this challenging task.

Human detection is the task of locating all instances of human beings present in an image, which has a wide range of applications across various fields, including search and rescue, surveillance, and autonomous driving. The rapid advancement of computer vision and deep learning technologies has brought significant improvements in human detection. However, for more advanced applications like healthcare, human–computer interaction, and scene understanding, it is crucial to obtain information beyond just the localization of humans. These applications require a deeper understanding of human behavior and state to enable effective and safe interactions with humans and the environment. This study presents a comprehensive benchmark, the Common Human Postures (CHP) dataset, aimed at promoting a more informative and more encouraging task beyond mere human detection. The benchmark dataset comprises a diverse collection of images, featuring individuals in different environments, clothing, and occlusions, performing a wide range of postures and activities. The benchmark aims to enhance research in this challenging task by designing novel and precise methods specifically for it. The CHP dataset consists of 5250 human images collected from different scenes, annotated with bounding boxes for seven common human poses. Using this well-annotated dataset, we have developed two baseline detectors, namely CHP-YOLOF and CHP-YOLOX, building upon two identity-preserved human posture detectors: IPH-YOLOF and IPH-YOLOX. We evaluate the performance of these baseline detectors through extensive experiments. The results demonstrate that these baseline detectors effectively detect human postures on the CHP dataset. By releasing the CHP dataset, we aim to facilitate further research on human pose estimation and to attract more researchers to focus on this challenging task.

Single-stream architectures using Vision Transformer (ViT) backbones show great potential for real-time UAV tracking recently. However, frequent occlusions from obstacles like buildings and trees expose a major drawback: these models often lack strategies to handle occlusions effectively. New methods are needed to enhance the occlusion resilience of single-stream ViT models in aerial tracking. In this work, we propose to learn Occlusion-Robust Representations (ORR) based on ViTs for UAV tracking by enforcing an invariance of the feature representation of a target with respect to random masking operations modeled by a spatial Cox process. Hopefully, this random masking approximately simulates target occlusions, thereby enabling us to learn ViTs that are robust to target occlusion for UAV tracking. This framework is termed ORTrack. Additionally, to facilitate real-time applications, we propose an Adaptive Feature-Based Knowledge Distillation (AFKD) method to create a more compact tracker, which adaptively mimics the behavior of the teacher model ORTrack according to the task's difficulty. This student model, dubbed ORTrack-D, retains much of ORTrack's performance while offering higher efficiency. Extensive experiments on multiple benchmarks validate the effectiveness of our method, demonstrating its state-of-the-art performance. Codes is available at https://github.com/wuyou3474/ORTrack.

Efficient UAV tracking faces challenges due to limited onboard resources, short battery life, and payload constraints. discriminative correlation filters (DCF)-based trackers have traditionally been preferred for their efficiency. However, recent lightweight deep learning (DL) trackers, leveraging model compression, exhibit significant CPU efficiency and precision. Unfortunately, existing model compression approaches have struggled with maintaining tracking precision, especially at higher compression rates. This paper addresses this limitation with a novel approach, called disentangled representation learning with mutual information maximization (DR-MIM). DR-MIM has the potential to enhance the accuracy and efficiency of DL-based trackers, particularly in UAV tracking. The primary concept underpinning this approach is disentangled representation which bifurcates the feature space into two distinct categories: identity-related and identity-unrelated elements. This bifurcation serves as a critical component, powering the heightened level of precision and efficiency attained by DL-based trackers. The focus of our work is predominantly on the employment of exclusively identity-related features. As a result of this focus, there’s a noticeable escalation in the efficacy of the feature representation. To further enhance efficiency without compromising accuracy, we have quantified the network, which has been named DR-MIM (v2). Exhaustive experiments conducted across four aerial tracking benchmarks have demonstrated the superiority of the proposed method. A standout performance on VisDrone2018 showcases DR-MIM (v2)’s impressive GPU speed of 658 FPS, coupled with a maintained precision of 82.1%, underscoring the potential of our approach in real-world UAV tracking scenarios.