Computer Science

While deep learning has reshaped the classical motion capture pipeline, generative, analysis-by-synthesis elements are still in use to recover fine details if a high-quality 3D model of the user is available. Unfortunately, obtaining such a model for every user a priori is challenging, time-consuming, and limits the application scenarios. We propose a novel test-time optimization approach for monocular motion capture that learns a volumetric body model of the user in a self-supervised manner. To this end, our approach combines the advantages of neural radiance fields with an articulated skeleton representation. Our proposed skeleton embedding serves as a common reference that links constraints across time, thereby reducing the number of required camera views from traditionally dozens of calibrated cameras, down to a single uncalibrated one. As a starting point, we employ the output of an off-the-shelf model that predicts the 3D skeleton pose. The volumetric body shape and appearance is then learned from scratch, while jointly refining the initial pose estimate. Our approach is self-supervised and does not require any additional ground truth labels for appearance, pose, or 3D shape. We demonstrate that our novel combination of a discriminative pose estimation technique with surface-free analysis-by-synthesis outperforms purely discriminative monocular pose estimation approaches and generalizes well to multiple views.

Semantic segmentation of remotely sensed images plays a crucial role in precision agriculture, environmental protection, and economic assessment. In recent years, substantial fine-resolution remote sensing images are available for semantic segmentation. However, due to the complicated information caused by the increased spatial resolution, state-of-the-art deep learning algorithms normally utilize complex network architectures for segmentation, which usually incurs high computational complexity. Specifically, the high-caliber performance of the convolutional neural network (CNN) heavily relies on fine-grained spatial details (fine resolution) and sufficient contextual information (large receptive fields), both of which trigger high computational costs. This crucially impedes their practicability and availability in real-world scenarios that require real-time processing. In this paper, we propose an Attentive Bilateral Contextual Network (ABCNet), a convolutional neural network (CNN) with double branches, with prominently lower computational consumptions compared to the cutting-edge algorithms, while maintaining a competitive accuracy. Code is available at this https URL.

Interpreting human actions requires understanding the spatial and temporal context of the scenes. State-of-the-art action detectors based on Convolutional Neural Network (CNN) have demonstrated remarkable results by adopting two-stream or 3D CNN architectures. However, these methods typically operate in a non-real-time, ofline fashion due to system complexity to reason spatio-temporal information. Consequently, their high computational cost is not compliant with emerging real-world scenarios such as service robots or public surveillance where detection needs to take place at resource-limited edge devices. In this paper, we propose ACDnet, a compact action detection network targeting real-time edge computing which addresses both efficiency and accuracy. It intelligently exploits the temporal coherence between successive video frames to approximate their CNN features rather than naively extracting them. It also integrates memory feature aggregation from past video frames to enhance current detection stability, implicitly modeling long temporal cues over time. Experiments conducted on the public benchmark datasets UCF-24 and JHMDB-21 demonstrate that ACDnet, when integrated with the SSD detector, can robustly achieve detection well above real-time (75 FPS). At the same time, it retains reasonable accuracy (70.92 and 49.53 frame mAP) compared to other top-performing methods using far heavier configurations. Codes will be available at this https URL.

Navigation inside a closed area with no GPS-signal accessibility is a highly challenging task. In order to tackle this problem, recently the imaging-based methods have grabbed the attention of many researchers. These methods either extract the features (e.g. using SIFT, or SOSNet) and map the descriptive ones to the camera position and rotation information, or deploy an end-to-end system that directly estimates this information out of RGB images, similar to PoseNet. While the former methods suffer from heavy computational burden during the test process, the latter suffers from lack of accuracy and robustness against environmental changes and object movements. However, end-to-end systems are quite fast during the test and inference and are pretty qualified for real-world applications, even though their training phase could be longer than the former ones. In this paper, a novel multi-modal end-to-end system for large-scale indoor positioning has been proposed, namely APS (Alpha Positioning System), which integrates a Pix2Pix GAN network to reconstruct the point cloud pair of the input query image, with a deep CNN network in order to robustly estimate the position and rotation information of the camera. For this integration, the existing datasets have the shortcoming of paired RGB/point cloud images for indoor environments. Therefore, we created a new dataset to handle this situation. By implementing the proposed APS system, we could achieve a highly accurate camera positioning with a precision level of less than a centimeter.

We propose a novel feature re-identification method for real-time visual-inertial SLAM. The front-end module of the state-of-the-art visual-inertial SLAM methods (e.g. visual feature extraction and matching schemes) relies on feature tracks across image frames, which are easily broken in challenging scenarios, resulting in insufficient visual measurement and accumulated error in pose estimation. In this paper, we propose an efficient drift-less SLAM method by re-identifying existing features from a spatial-temporal sensitive sub-global map. The re-identified features over a long time span serve as augmented visual measurements and are incorporated into the optimization module which can gradually decrease the accumulative error in the long run, and further build a drift-less global map in the system. Extensive experiments show that our feature re-identification method is both effective and efficient. Specifically, when combining the feature re-identification with the state-of-the-art SLAM method [11], our method achieves 67.3% and 87.5% absolute translation error reduction with only a small additional computational cost on two public SLAM benchmark DBs: EuRoC and TUM-VI respectively.

Precipitation nowcasting is an important task for weather forecasting. Many recent works aim to predict the high rainfall events more accurately with the help of deep learning techniques, but such events are relatively rare. The rarity is often addressed by formulations that re-weight the rare events. Somehow such a formulation carries a side effect of making "blurry" predictions in low rainfall regions and cannot convince meteorologists to trust its practical usability. We fix the trust issue by introducing a discriminator that encourages the prediction model to generate realistic rain-maps without sacrificing predictive accuracy. Furthermore, we extend the nowcasting time frame from one hour to three hours to further address the needs from meteorologists. The extension is based on consecutive attentions across different hours. We propose a new deep learning model for precipitation nowcasting that includes both the discrimination and attention techniques. The model is examined on a newly-built benchmark dataset that contains both radar data and actual rain data. The benchmark, which will be publicly released, not only establishes the superiority of the proposed model, but also is expected to encourage future research on precipitation nowcasting.

Agricultural image recognition tasks are becoming increasingly dependent on deep learning (DL). Despite its excellent performance, it is difficult to comprehend what type of logic or features DL uses in its decision making. This has become a roadblock for the implementation and development of DL-based image recognition methods because knowing the logic or features used in decision making, such as in a classification task, is very important for verification, algorithm improvement, training data improvement, knowledge extraction, etc. To mitigate such problems, we developed a classification method based on a variational autoencoder architecture that can show not only the location of the most important features but also what variations of that particular feature are used. Using the PlantVillage dataset, we achieved an acceptable level of explainability without sacrificing the accuracy of the classification. Although the proposed method was tested for disease diagnosis in some crops, the method can be extended to other crops as well as other image classification tasks. In the future, we hope to use this explainable artificial intelligence algorithm in disease identification tasks, such as the identification of potato blackleg disease and potato virus Y (PVY), and other image classification tasks.

This paper proposes a novel active boundary loss for semantic segmentation. It can progressively encourage the alignment between predicted boundaries and ground-truth boundaries during end-to-end training, which is not explicitly enforced in commonly used cross-entropy loss. Based on the predicted boundaries detected from the segmentation results using current network parameters, we formulate the boundary alignment problem as a differentiable direction vector prediction problem to guide the movement of predicted boundaries in each iteration. Our loss is model-agnostic and can be plugged into the training of segmentation networks to improve the boundary details. Experimental results show that training with the active boundary loss can effectively improve the boundary F-score and mean Intersection-over-Union on challenging image and video object segmentation datasets.

This paper describes a system by which Unmanned Aerial Vehicles (UAVs) can gather high-quality face images that can be used in biometric identification tasks. Success in face-based identification depends in large part on the image quality, and a major factor is how frontal the view is. Face recognition software pipelines can improve identification rates by synthesizing frontal views from non-frontal views by a process call {\em frontalization}. Here we exploit the high mobility of UAVs to actively gather frontal images using components of a synthetic frontalization pipeline. We define a frontalization error and show that it can be used to guide an UAVs to capture frontal views. Further, we show that the resulting image stream improves matching quality of a typical face recognition similarity metric. The system is implemented using an off-the-shelf hardware and software components and can be easily transfered to any ROS enabled UAVs.

Explainable AI (XAI) is an active research area to interpret a neural network's decision by ensuring transparency and trust in the task-specified learned models. Recently, perturbation-based model analysis has shown better interpretation, but backpropagation techniques are still prevailing because of their computational efficiency. In this work, we combine both approaches as a hybrid visual explanation algorithm and propose an efficient interpretation method for convolutional neural networks. Our method adaptively selects the most critical features that mainly contribute towards a prediction to probe the model by finding the activated features. Experimental results show that the proposed method can reduce the execution time up to 30% while enhancing competitive interpretability without compromising the quality of explanation generated.