Vision Transformers have enabled recent attention-based Deep Learning (DL) architectures to achieve remarkable results in Computer Vision (CV) tasks. However, due to the extensive computational resources required, these architectures are rarely implemented on resource-constrained platforms. Current research investigates hybrid handcrafted convolution-based and attention-based models for CV tasks such as image classification and object detection. In this paper, we propose HyT-NAS, an efficient Hardware-aware Neural Architecture Search (HW-NAS) including hybrid architectures targeting vision tasks on tiny devices. HyT-NAS improves state-of-the-art HW-NAS by enriching the search space and enhancing the search strategy as well as the performance predictors. Our experiments show that HyT-NAS achieves a similar hypervolume with less than ~5x training evaluations. Our resulting architecture outperforms MLPerf MobileNetV1 by 6.3% accuracy improvement with 3.5x less number of parameters on Visual Wake Words.
In recent years, deep learning has been successfully applied in various scientific domains. Following these promising results and performances, it has recently also started being evaluated in the domain of radio astronomy. In particular, since radio astronomy is entering the Big Data era, with the advent of the largest telescope in the world - the Square Kilometre Array (SKA), the task of automatic object detection and instance segmentation is crucial for source finding and analysis. In this work, we explore the performance of the most affirmed deep learning approaches, applied to astronomical images obtained by radio interferometric instrumentation, to solve the task of automatic source detection. This is carried out by applying models designed to accomplish two different kinds of tasks: object detection and semantic segmentation. The goal is to provide an overview of existing techniques, in terms of prediction performance and computational efficiency, to scientists in the astrophysics community who would like to employ machine learning in their research.
Point cloud processing methods exploit local point features and global context through aggregation which does not explicity model the internal correlations between local and global features. To address this problem, we propose full point encoding which is applicable to convolution and transformer architectures. Specifically, we propose Full Point Convolution (FPConv) and Full Point Transformer (FPTransformer) architectures. The key idea is to adaptively learn the weights from local and global geometric connections, where the connections are established through local and global correlation functions respectively. FPConv and FPTransformer simultaneously model the local and global geometric relationships as well as their internal correlations, demonstrating strong generalization ability and high performance. FPConv is incorporated in classical hierarchical network architectures to achieve local and global shape-aware learning. In FPTransformer, we introduce full point position encoding in self-attention, that hierarchically encodes each point position in the global and local receptive field. We also propose a shape aware downsampling block which takes into account the local shape and the global context. Experimental comparison to existing methods on benchmark datasets show the efficacy of FPConv and FPTransformer for semantic segmentation, object detection, classification, and normal estimation tasks. In particular, we achieve state-of-the-art semantic segmentation results of 76% mIoU on S3DIS 6-fold and 72.2% on S3DIS Area5.
Adversarial attacks on deep-learning models have been receiving increased attention in recent years. Work in this area has mostly focused on gradient-based techniques, so-called "white-box" attacks, wherein the attacker has access to the targeted model's internal parameters; such an assumption is usually unrealistic in the real world. Some attacks additionally use the entire pixel space to fool a given model, which is neither practical nor physical (i.e., real-world). On the contrary, we propose herein a direct, black-box, gradient-free method that uses the learned image manifold of a pretrained generative adversarial network (GAN) to generate naturalistic physical adversarial patches for object detectors. To our knowledge this is the first and only method that performs black-box physical attacks directly on object-detection models, which results with a model-agnostic attack. We show that our proposed method works both digitally and physically. We compared our approach against four different black-box attacks with different configurations. Our approach outperformed all other approaches that were tested in our experiments by a large margin.
Domain adaptive object detection (DAOD) assumes that both labeled source data and unlabeled target data are available for training, but this assumption does not always hold in real-world scenarios. Thus, source-free DAOD is proposed to adapt the source-trained detectors to target domains with only unlabeled target data. Existing source-free DAOD methods typically utilize pseudo labeling, where the performance heavily relies on the selection of confidence threshold. However, most prior works adopt a single fixed threshold for all classes to generate pseudo labels, which ignore the imbalanced class distribution, resulting in biased pseudo labels. In this work, we propose a refined pseudo labeling framework for source-free DAOD. First, to generate unbiased pseudo labels, we present a category-aware adaptive threshold estimation module, which adaptively provides the appropriate threshold for each category. Second, to alleviate incorrect box regression, a localization-aware pseudo label assignment strategy is introduced to divide labels into certain and uncertain ones and optimize them separately. Finally, extensive experiments on four adaptation tasks demonstrate the effectiveness of our method.
Domain adaptive object detection (DAOD) aims to adapt the detector from a labelled source domain to an unlabelled target domain. In recent years, DAOD has attracted massive attention since it can alleviate performance degradation due to the large shift of data distributions in the wild. To align distributions between domains, adversarial learning is widely used in existing DAOD methods. However, the decision boundary for the adversarial domain discriminator may be inaccurate, causing the model biased towards the source domain. To alleviate this bias, we propose a novel Frequency-based Image Translation (FIT) framework for DAOD. First, by keeping domain-invariant frequency components and swapping domain-specific ones, we conduct image translation to reduce domain shift at the input level. Second, hierarchical adversarial feature learning is utilized to further mitigate the domain gap at the feature level. Finally, we design a joint loss to train the entire network in an end-to-end manner without extra training to obtain translated images. Extensive experiments on three challenging DAOD benchmarks demonstrate the effectiveness of our method.
LiDAR-camera fusion methods have shown impressive performance in 3D object detection. Recent advanced multi-modal methods mainly perform global fusion, where image features and point cloud features are fused across the whole scene. Such practice lacks fine-grained region-level information, yielding suboptimal fusion performance. In this paper, we present the novel Local-to-Global fusion network (LoGoNet), which performs LiDAR-camera fusion at both local and global levels. Concretely, the Global Fusion (GoF) of LoGoNet is built upon previous literature, while we exclusively use point centroids to more precisely represent the position of voxel features, thus achieving better cross-modal alignment. As to the Local Fusion (LoF), we first divide each proposal into uniform grids and then project these grid centers to the images. The image features around the projected grid points are sampled to be fused with position-decorated point cloud features, maximally utilizing the rich contextual information around the proposals. The Feature Dynamic Aggregation (FDA) module is further proposed to achieve information interaction between these locally and globally fused features, thus producing more informative multi-modal features. Extensive experiments on both Waymo Open Dataset (WOD) and KITTI datasets show that LoGoNet outperforms all state-of-the-art 3D detection methods. Notably, LoGoNet ranks 1st on Waymo 3D object detection leaderboard and obtains 81.02 mAPH (L2) detection performance. It is noteworthy that, for the first time, the detection performance on three classes surpasses 80 APH (L2) simultaneously. Code will be available at \url{https://github.com/sankin97/LoGoNet}.
Effective BEV object detection on infrastructure can greatly improve traffic scenes understanding and vehicle-toinfrastructure (V2I) cooperative perception. However, cameras installed on infrastructure have various postures, and previous BEV detection methods rely on accurate calibration, which is difficult for practical applications due to inevitable natural factors (e.g., wind and snow). In this paper, we propose a Calibration-free BEV Representation (CBR) network, which achieves 3D detection based on BEV representation without calibration parameters and additional depth supervision. Specifically, we utilize two multi-layer perceptrons for decoupling the features from perspective view to front view and birdeye view under boxes-induced foreground supervision. Then, a cross-view feature fusion module matches features from orthogonal views according to similarity and conducts BEV feature enhancement with front view features. Experimental results on DAIR-V2X demonstrate that CBR achieves acceptable performance without any camera parameters and is naturally not affected by calibration noises. We hope CBR can serve as a baseline for future research addressing practical challenges of infrastructure perception.
This paper introduces a novel approach to video object detection detection and tracking on Unmanned Aerial Vehicles (UAVs). By incorporating metadata, the proposed approach creates a memory map of object locations in actual world coordinates, providing a more robust and interpretable representation of object locations in both, image space and the real world. We use this representation to boost confidences, resulting in improved performance for several temporal computer vision tasks, such as video object detection, short and long-term single and multi-object tracking, and video anomaly detection. These findings confirm the benefits of metadata in enhancing the capabilities of UAVs in the field of temporal computer vision and pave the way for further advancements in this area.
Multi-head-self-attention (MHSA)-equipped models have achieved notable performance in computer vision. Their computational complexity is proportional to quadratic numbers of pixels in input feature maps, resulting in slow processing, especially when dealing with high-resolution images. New types of token-mixer are proposed as an alternative to MHSA to circumvent this problem: an FFT-based token-mixer involves global operations similar to MHSA but with lower computational complexity. However, despite its attractive properties, the FFT-based token-mixer has not been carefully examined in terms of its compatibility with the rapidly evolving MetaFormer architecture. Here, we propose a novel token-mixer called Dynamic Filter and novel image recognition models, DFFormer and CDFFormer, to close the gaps above. The results of image classification and downstream tasks, analysis, and visualization show that our models are helpful. Notably, their throughput and memory efficiency when dealing with high-resolution image recognition is remarkable. Our results indicate that Dynamic Filter is one of the token-mixer options that should be seriously considered. The code is available at https://github.com/okojoalg/dfformer
We present LGX (Language-guided Exploration), a novel algorithm for Language-Driven Zero-Shot Object Goal Navigation (L-ZSON), where an embodied agent navigates to a uniquely described target object in a previously unseen environment. Our approach makes use of Large Language Models (LLMs) for this task by leveraging the LLM's commonsense reasoning capabilities for making sequential navigational decisions. Simultaneously, we perform generalized target object detection using a pre-trained Vision-Language grounding model. We achieve state-of-the-art zero-shot object navigation results on RoboTHOR with a success rate (SR) improvement of over 27% over the current baseline of the OWL-ViT CLIP on Wheels (OWL CoW). Furthermore, we study the usage of LLMs for robot navigation and present an analysis of various prompting strategies affecting the model output. Finally, we showcase the benefits of our approach via \textit{real-world} experiments that indicate the superior performance of LGX in detecting and navigating to visually unique objects.
Recently, virtual/pseudo-point-based 3D object detection that seamlessly fuses RGB images and LiDAR data by depth completion has gained great attention. However, virtual points generated from an image are very dense, introducing a huge amount of redundant computation during detection. Meanwhile, noises brought by inaccurate depth completion significantly degrade detection precision. This paper proposes a fast yet effective backbone, termed VirConvNet, based on a new operator VirConv (Virtual Sparse Convolution), for virtual-point-based 3D object detection. VirConv consists of two key designs: (1) StVD (Stochastic Voxel Discard) and (2) NRConv (Noise-Resistant Submanifold Convolution). StVD alleviates the computation problem by discarding large amounts of nearby redundant voxels. NRConv tackles the noise problem by encoding voxel features in both 2D image and 3D LiDAR space. By integrating VirConv, we first develop an efficient pipeline VirConv-L based on an early fusion design. Then, we build a high-precision pipeline VirConv-T based on a transformed refinement scheme. Finally, we develop a semi-supervised pipeline VirConv-S based on a pseudo-label framework. On the KITTI car 3D detection test leaderboard, our VirConv-L achieves 85% AP with a fast running speed of 56ms. Our VirConv-T and VirConv-S attains a high-precision of 86.3% and 87.2% AP, and currently rank 2nd and 1st, respectively. The code is available at https://github.com/hailanyi/VirConv.
The recent surge in interest in autonomous driving stems from its rapidly developing capacity to enhance safety, efficiency, and convenience. A pivotal aspect of autonomous driving technology is its perceptual systems, where core algorithms have yielded more precise algorithms applicable to autonomous driving, including vision-based Simultaneous Localization and Mapping (SLAMs), object detection, and tracking algorithms. This work introduces a visual-based perception system for autonomous driving that integrates trajectory tracking and prediction of moving objects to prevent collisions, while addressing autonomous driving's localization and mapping requirements. The system leverages motion cues from pedestrians to monitor and forecast their movements and simultaneously maps the environment. This integrated approach resolves camera localization and the tracking of other moving objects in the scene, subsequently generating a sparse map to facilitate vehicle navigation. The performance, efficiency, and resilience of this approach are substantiated through comprehensive evaluations of both simulated and real-world datasets.
Recent advances in 3D object detection (3DOD) have obtained remarkably strong results for LiDAR-based models. In contrast, surround-view 3DOD models based on multiple camera images underperform due to the necessary view transformation of features from perspective view (PV) to a 3D world representation which is ambiguous due to missing depth information. This paper introduces X$^3$KD, a comprehensive knowledge distillation framework across different modalities, tasks, and stages for multi-camera 3DOD. Specifically, we propose cross-task distillation from an instance segmentation teacher (X-IS) in the PV feature extraction stage providing supervision without ambiguous error backpropagation through the view transformation. After the transformation, we apply cross-modal feature distillation (X-FD) and adversarial training (X-AT) to improve the 3D world representation of multi-camera features through the information contained in a LiDAR-based 3DOD teacher. Finally, we also employ this teacher for cross-modal output distillation (X-OD), providing dense supervision at the prediction stage. We perform extensive ablations of knowledge distillation at different stages of multi-camera 3DOD. Our final X$^3$KD model outperforms previous state-of-the-art approaches on the nuScenes and Waymo datasets and generalizes to RADAR-based 3DOD. Qualitative results video at https://youtu.be/1do9DPFmr38.
Pseudo-Labeling has emerged as a simple yet effective technique for semi-supervised object detection (SSOD). However, the inevitable noise problem in pseudo-labels significantly degrades the performance of SSOD methods. Recent advances effectively alleviate the classification noise in SSOD, while the localization noise which is a non-negligible part of SSOD is not well-addressed. In this paper, we analyse the localization noise from the generation and learning phases, and propose two strategies, namely pseudo-label correction and noise-unaware learning. For pseudo-label correction, we introduce a multi-round refining method and a multi-vote weighting method. The former iteratively refines the pseudo boxes to improve the stability of predictions, while the latter smoothly self-corrects pseudo boxes by weighing the scores of surrounding jittered boxes. For noise-unaware learning, we introduce a loss weight function that is negatively correlated with the Intersection over Union (IoU) in the regression task, which pulls the predicted boxes closer to the object and improves localization accuracy. Our proposed method, Pseudo-label Correction and Learning (PCL), is extensively evaluated on the MS COCO and PASCAL VOC benchmarks. On MS COCO, PCL outperforms the supervised baseline by 12.16, 12.11, and 9.57 mAP and the recent SOTA (SoftTeacher) by 3.90, 2.54, and 2.43 mAP under 1\%, 5\%, and 10\% labeling ratios, respectively. On PASCAL VOC, PCL improves the supervised baseline by 5.64 mAP and the recent SOTA (Unbiased Teacherv2) by 1.04 mAP on AP$^{50}$.
Human-centric perceptions (e.g., pose estimation, human parsing, pedestrian detection, person re-identification, etc.) play a key role in industrial applications of visual models. While specific human-centric tasks have their own relevant semantic aspect to focus on, they also share the same underlying semantic structure of the human body. However, few works have attempted to exploit such homogeneity and design a general-propose model for human-centric tasks. In this work, we revisit a broad range of human-centric tasks and unify them in a minimalist manner. We propose UniHCP, a Unified Model for Human-Centric Perceptions, which unifies a wide range of human-centric tasks in a simplified end-to-end manner with the plain vision transformer architecture. With large-scale joint training on 33 human-centric datasets, UniHCP can outperform strong baselines on several in-domain and downstream tasks by direct evaluation. When adapted to a specific task, UniHCP achieves new SOTAs on a wide range of human-centric tasks, e.g., 69.8 mIoU on CIHP for human parsing, 86.18 mA on PA-100K for attribute prediction, 90.3 mAP on Market1501 for ReID, and 85.8 JI on CrowdHuman for pedestrian detection, performing better than specialized models tailored for each task.
Inspired by the recent success of application of dense data approach by using ORB-SLAM and RGB-D SLAM, we propose a better pipeline of real-time SLAM in dynamics environment. Different from previous SLAM which can only handle static scenes, we are presenting a solution which use RGB-D SLAM as well as YOLO real-time object detection to segment and remove dynamic scene and then construct static scene 3D. We gathered a dataset which allows us to jointly consider semantics, geometry, and physics and thus enables us to reconstruct the static scene while filtering out all dynamic objects.
Object detectors used in autonomous vehicles can have high memory and computational overheads. In this paper, we introduce a novel semi-structured pruning framework called R-TOSS that overcomes the shortcomings of state-of-the-art model pruning techniques. Experimental results on the JetsonTX2 show that R-TOSS has a compression rate of 4.4x on the YOLOv5 object detector with a 2.15x speedup in inference time and 57.01% decrease in energy usage. R-TOSS also enables 2.89x compression on RetinaNet with a 1.86x speedup in inference time and 56.31% decrease in energy usage. We also demonstrate significant improvements compared to various state-of-the-art pruning techniques.
Searching for objects is a fundamental skill for robots. As such, we expect object search to eventually become an off-the-shelf capability for robots, similar to e.g., object detection and SLAM. In contrast, however, no system for 3D object search exists that generalizes across real robots and environments. In this paper, building upon a recent theoretical framework that exploited the octree structure for representing belief in 3D, we present GenMOS (Generalized Multi-Object Search), the first general-purpose system for multi-object search (MOS) in a 3D region that is robot-independent and environment-agnostic. GenMOS takes as input point cloud observations of the local region, object detection results, and localization of the robot's view pose, and outputs a 6D viewpoint to move to through online planning. In particular, GenMOS uses point cloud observations in three ways: (1) to simulate occlusion; (2) to inform occupancy and initialize octree belief; and (3) to sample a belief-dependent graph of view positions that avoid obstacles. We evaluate our system both in simulation and on two real robot platforms. Our system enables, for example, a Boston Dynamics Spot robot to find a toy cat hidden underneath a couch in under one minute. We further integrate 3D local search with 2D global search to handle larger areas, demonstrating the resulting system in a 25m$^2$ lobby area.
The use of ground control points (GCPs) for georeferencing is the most common strategy in unmanned aerial vehicle (UAV) photogrammetry, but at the same time their collection represents the most time-consuming and expensive part of UAV campaigns. Recently, deep learning has been rapidly developed in the field of small object detection. In this letter, to automatically extract coordinates information of ground control points (GCPs) by detecting GCP-markers in UAV images, we propose a solution that uses a deep learning-based architecture, YOLOv5-OBB, combined with a confidence threshold filtering algorithm and an optimal ranking algorithm. We applied our proposed method to a dataset collected by DJI Phantom 4 Pro drone and obtained good detection performance with the mean Average Precision (AP) of 0.832 and the highest AP of 0.982 for the cross-type GCP-markers. The proposed method can be a promising tool for future implementation of the end-to-end aerial triangulation process.