Collecting and annotating real-world data for the development of object detection models is a time-consuming and expensive process. In the military domain in particular, data collection can also be dangerous or infeasible. Training models on synthetic data may provide a solution for cases where access to real-world training data is restricted. However, bridging the reality gap between synthetic and real data remains a challenge. Existing methods usually build on top of baseline Convolutional Neural Network (CNN) models that have been shown to perform well when trained on real data, but have limited ability to perform well when trained on synthetic data. For example, some architectures allow for fine-tuning with the expectation of large quantities of training data and are prone to overfitting on synthetic data. Related work usually ignores various best practices from object detection on real data, e.g. by training on synthetic data from a single environment with relatively little variation. In this paper we propose a methodology for improving the performance of a pre-trained object detector when training on synthetic data. Our approach focuses on extracting the salient information from synthetic data without forgetting useful features learned from pre-training on real images. Based on the state of the art, we incorporate data augmentation methods and a Transformer backbone. Besides reaching relatively strong performance without any specialized synthetic data transfer methods, we show that our methods improve the state of the art on synthetic data trained object detection for the RarePlanes and DGTA-VisDrone datasets, and reach near-perfect performance on an in-house vehicle detection dataset.
Deep neural networks (DNNs) have been proven extremely susceptible to adversarial examples, which raises special safety-critical concerns for DNN-based autonomous driving stacks (i.e., 3D object detection). Although there are extensive works on image-level attacks, most are restricted to 2D pixel spaces, and such attacks are not always physically realistic in our 3D world. Here we present Adv3D, the first exploration of modeling adversarial examples as Neural Radiance Fields (NeRFs). Advances in NeRF provide photorealistic appearances and 3D accurate generation, yielding a more realistic and realizable adversarial example. We train our adversarial NeRF by minimizing the surrounding objects' confidence predicted by 3D detectors on the training set. Then we evaluate Adv3D on the unseen validation set and show that it can cause a large performance reduction when rendering NeRF in any sampled pose. To generate physically realizable adversarial examples, we propose primitive-aware sampling and semantic-guided regularization that enable 3D patch attacks with camouflage adversarial texture. Experimental results demonstrate that the trained adversarial NeRF generalizes well to different poses, scenes, and 3D detectors. Finally, we provide a defense method to our attacks that involves adversarial training through data augmentation. Project page: https://len-li.github.io/adv3d-web
Object detection forms a key component in Unmanned Aerial Vehicles (UAVs) for completing high-level tasks that depend on the awareness of objects on the ground from an aerial perspective. In that scenario, adversarial patch attacks on an onboard object detector can severely impair the performance of upstream tasks. This paper proposes a novel model-agnostic defense mechanism against the threat of adversarial patch attacks in the context of UAV-based object detection. We formulate adversarial patch defense as an occlusion removal task. The proposed defense method can neutralize adversarial patches located on objects of interest, without exposure to adversarial patches during training. Our lightweight single-stage defense approach allows us to maintain a model-agnostic nature, that once deployed does not require to be updated in response to changes in the object detection pipeline. The evaluations in digital and physical domains show the feasibility of our method for deployment in UAV object detection pipelines, by significantly decreasing the Attack Success Ratio without incurring significant processing costs. As a result, the proposed defense solution can improve the reliability of object detection for UAVs.
Multispectral object detection, utilizing both visible (RGB) and thermal infrared (T) modals, has garnered significant attention for its robust performance across diverse weather and lighting conditions. However, effectively exploiting the complementarity between RGB-T modals while maintaining efficiency remains a critical challenge. In this paper, a very simple Group Shuffled Multi-receptive Attention (GSMA) module is proposed to extract and combine multi-scale RGB and thermal features. Then, the extracted multi-modal features are directly integrated with a multi-level path aggregation neck, which significantly improves the fusion effect and efficiency. Meanwhile, multi-modal object detection often adopts union annotations for both modals. This kind of supervision is not sufficient and unfair, since objects observed in one modal may not be seen in the other modal. To solve this issue, Multi-modal Supervision (MS) is proposed to sufficiently supervise RGB-T object detection. Comprehensive experiments on two challenging benchmarks, KAIST and DroneVehicle, demonstrate the proposed model achieves the state-of-the-art accuracy while maintaining competitive efficiency.
Visual-based perception is the key module for autonomous driving. Among those visual perception tasks, video object detection is a primary yet challenging one because of feature degradation caused by fast motion or multiple poses. Current models usually aggregate features from the neighboring frames to enhance the object representations for the task heads to generate more accurate predictions. Though getting better performance, these methods rely on the information from the future frames and suffer from high computational complexity. Meanwhile, the aggregation process is not reconfigurable during the inference time. These issues make most of the existing models infeasible for online applications. To solve these problems, we introduce a stepwise spatial global-local aggregation network. Our proposed models mainly contain three parts: 1). Multi-stage stepwise network gradually refines the predictions and object representations from the previous stage; 2). Spatial global-local aggregation fuses the local information from the neighboring frames and global semantics from the current frame to eliminate the feature degradation; 3). Dynamic aggregation strategy stops the aggregation process early based on the refinement results to remove redundancy and improve efficiency. Extensive experiments on the ImageNet VID benchmark validate the effectiveness and efficiency of our proposed models.
Real-time and high-performance 3D object detection plays a critical role in autonomous driving and robotics. Recent pillar-based 3D object detectors have gained significant attention due to their compact representation and low computational overhead, making them suitable for onboard deployment and quantization. However, existing pillar-based detectors still suffer from information loss along height dimension and large numerical distribution difference during pillar feature encoding (PFE), which severely limits their performance and quantization potential. To address above issue, we first unveil the importance of different input information during PFE and identify the height dimension as a key factor in enhancing 3D detection performance. Motivated by this observation, we propose a height-aware pillar feature encoder, called PillarHist. Specifically, PillarHist statistics the discrete distribution of points at different heights within one pillar with the information entropy guidance. This simple yet effective design greatly preserves the information along the height dimension while significantly reducing the computation overhead of the PFE. Meanwhile, PillarHist also constrains the arithmetic distribution of PFE input to a stable range, making it quantization-friendly. Notably, PillarHist operates exclusively within the PFE stage to enhance performance, enabling seamless integration into existing pillar-based methods without introducing complex operations. Extensive experiments show the effectiveness of PillarHist in terms of both efficiency and performance.
Deep learning-based information processing consumes long time and requires huge computing resources, especially for dense prediction tasks which require an output for each pixel, like semantic segmentation and salient object detection. There are mainly two challenges for quantization of dense prediction tasks. Firstly, directly applying the upsampling operation that dense prediction tasks require is extremely crude and causes unacceptable accuracy reduction. Secondly, the complex structure of dense prediction networks means it is difficult to maintain a fast speed as well as a high accuracy when performing quantization. In this paper, we propose an effective upsampling method and an efficient attention computation strategy to transfer the success of the binary neural networks (BNN) from single prediction tasks to dense prediction tasks. Firstly, we design a simple and robust multi-branch parallel upsampling structure to achieve the high accuracy. Then we further optimize the attention method which plays an important role in segmentation but has huge computation complexity. Our attention method can reduce the computational complexity by a factor of one hundred times but retain the original effect. Experiments on Cityscapes, KITTI road, and ECSSD fully show the effectiveness of our work.
Transformer-based detectors have shown success in computer vision tasks with natural images. These models, exemplified by the Deformable DETR, are optimized through complex engineering strategies tailored to the typical characteristics of natural scenes. However, medical imaging data presents unique challenges such as extremely large image sizes, fewer and smaller regions of interest, and object classes which can be differentiated only through subtle differences. This study evaluates the applicability of these transformer-based design choices when applied to a screening mammography dataset that represents these distinct medical imaging data characteristics. Our analysis reveals that common design choices from the natural image domain, such as complex encoder architectures, multi-scale feature fusion, query initialization, and iterative bounding box refinement, do not improve and sometimes even impair object detection performance in medical imaging. In contrast, simpler and shallower architectures often achieve equal or superior results. This finding suggests that the adaptation of transformer models for medical imaging data requires a reevaluation of standard practices, potentially leading to more efficient and specialized frameworks for medical diagnosis.
Image matting aims to obtain an alpha matte that separates foreground objects from the background accurately. Recently, trimap-free matting has been well studied because it requires only the original image without any extra input. Such methods usually extract a rough foreground by itself to take place trimap as further guidance. However, the definition of 'foreground' lacks a unified standard and thus ambiguities arise. Besides, the extracted foreground is sometimes incomplete due to inadequate network design. Most importantly, there is not a large-scale real-world matting dataset, and current trimap-free methods trained with synthetic images suffer from large domain shift problems in practice. In this paper, we define the salient object as foreground, which is consistent with human cognition and annotations of the current matting dataset. Meanwhile, data and technologies in salient object detection can be transferred to matting in a breeze. To obtain a more accurate and complete alpha matte, we propose a network called \textbf{M}ulti-\textbf{F}eature fusion-based \textbf{C}oarse-to-fine Network \textbf{(MFC-Net)}, which fully integrates multiple features for an accurate and complete alpha matte. Furthermore, we introduce image harmony in data composition to bridge the gap between synthetic and real images. More importantly, we establish the largest general matting dataset \textbf{(Real-19k)} in the real world to date. Experiments show that our method is significantly effective on both synthetic and real-world images, and the performance in the real-world dataset is far better than existing matting-free methods. Our code and data will be released soon.
Rapid advancements in Autonomous Driving (AD) tasks turned a significant shift toward end-to-end fashion, particularly in the utilization of vision-language models (VLMs) that integrate robust logical reasoning and cognitive abilities to enable comprehensive end-to-end planning. However, these VLM-based approaches tend to integrate 2D vision tokenizers and a large language model (LLM) for ego-car planning, which lack 3D geometric priors as a cornerstone of reliable planning. Naturally, this observation raises a critical concern: Can a 2D-tokenized LLM accurately perceive the 3D environment? Our evaluation of current VLM-based methods across 3D object detection, vectorized map construction, and environmental caption suggests that the answer is, unfortunately, NO. In other words, 2D-tokenized LLM fails to provide reliable autonomous driving. In response, we introduce DETR-style 3D perceptrons as 3D tokenizers, which connect LLM with a one-layer linear projector. This simple yet elegant strategy, termed Atlas, harnesses the inherent priors of the 3D physical world, enabling it to simultaneously process high-resolution multi-view images and employ spatiotemporal modeling. Despite its simplicity, Atlas demonstrates superior performance in both 3D detection and ego planning tasks on nuScenes dataset, proving that 3D-tokenized LLM is the key to reliable autonomous driving. The code and datasets will be released.
In real-life scenarios, humans seek out objects in the 3D world to fulfill their daily needs or intentions. This inspires us to introduce 3D intention grounding, a new task in 3D object detection employing RGB-D, based on human intention, such as "I want something to support my back". Closely related, 3D visual grounding focuses on understanding human reference. To achieve detection based on human intention, it relies on humans to observe the scene, reason out the target that aligns with their intention ("pillow" in this case), and finally provide a reference to the AI system, such as "A pillow on the couch". Instead, 3D intention grounding challenges AI agents to automatically observe, reason and detect the desired target solely based on human intention. To tackle this challenge, we introduce the new Intent3D dataset, consisting of 44,990 intention texts associated with 209 fine-grained classes from 1,042 scenes of the ScanNet dataset. We also establish several baselines based on different language-based 3D object detection models on our benchmark. Finally, we propose IntentNet, our unique approach, designed to tackle this intention-based detection problem. It focuses on three key aspects: intention understanding, reasoning to identify object candidates, and cascaded adaptive learning that leverages the intrinsic priority logic of different losses for multiple objective optimization.
Constructing large-scale labeled datasets for multi-modal perception model training in autonomous driving presents significant challenges. This has motivated the development of self-supervised pretraining strategies. However, existing pretraining methods mainly employ distinct approaches for each modality. In contrast, we focus on LiDAR-Camera 3D perception models and introduce a unified pretraining strategy, NeRF-Supervised Masked Auto Encoder (NS-MAE), which optimizes all modalities through a shared formulation. NS-MAE leverages NeRF's ability to encode both appearance and geometry, enabling efficient masked reconstruction of multi-modal data. Specifically, embeddings are extracted from corrupted LiDAR point clouds and images, conditioned on view directions and locations. Then, these embeddings are rendered into multi-modal feature maps from two crucial viewpoints for 3D driving perception: perspective and bird's-eye views. The original uncorrupted data serve as reconstruction targets for self-supervised learning. Extensive experiments demonstrate the superior transferability of NS-MAE across various 3D perception tasks under different fine-tuning settings. Notably, NS-MAE outperforms prior SOTA pre-training methods that employ separate strategies for each modality in BEV map segmentation under the label-efficient fine-tuning setting. Our code is publicly available at https://github.com/Xiaohao-Xu/Unified-Pretrain-AD/ .
Open-vocabulary detection aims to detect objects from novel categories beyond the base categories on which the detector is trained. However, existing open-vocabulary detectors trained on base category data tend to assign higher confidence to trained categories and confuse novel categories with the background. To resolve this, we propose OV-DQUO, an \textbf{O}pen-\textbf{V}ocabulary DETR with \textbf{D}enoising text \textbf{Q}uery training and open-world \textbf{U}nknown \textbf{O}bjects supervision. Specifically, we introduce a wildcard matching method. This method enables the detector to learn from pairs of unknown objects recognized by the open-world detector and text embeddings with general semantics, mitigating the confidence bias between base and novel categories. Additionally, we propose a denoising text query training strategy. It synthesizes foreground and background query-box pairs from open-world unknown objects to train the detector through contrastive learning, enhancing its ability to distinguish novel objects from the background. We conducted extensive experiments on the challenging OV-COCO and OV-LVIS benchmarks, achieving new state-of-the-art results of 45.6 AP50 and 39.3 mAP on novel categories respectively, without the need for additional training data. Models and code are released at \url{https://github.com/xiaomoguhz/OV-DQUO}
Convolutional Neural Networks (CNN) are commonly used for the problem of object detection thanks to their increased accuracy. Nevertheless, the performance of CNN-based detection models is ambiguous when detection speed is considered. To the best of our knowledge, there has not been sufficient evaluation of the available methods in terms of the speed/accuracy trade-off in related literature. This work assesses the most fundamental object detection models on the Common Objects in Context (COCO) dataset with respect to this trade-off, their memory consumption, and computational and storage cost. Next, we select a highly efficient model called YOLOv5 to train on the topical and unexplored dataset of human faces with medical masks, the Properly-Wearing Masked Faces Dataset (PWMFD), and analyze the benefits of specific optimization techniques for real-time medical mask detection: transfer learning, data augmentations, and a Squeeze-and-Excitation attention mechanism. Using our findings in the context of the COVID-19 pandemic, we propose an optimized model based on YOLOv5s using transfer learning for the detection of correctly and incorrectly worn medical masks that surpassed more than two times in speed (69 frames per second) the state-of-the-art model SE-YOLOv3 on the PWMFD dataset while maintaining the same level of mean Average Precision (67%).
The joint-embedding predictive architecture (JEPA) recently has shown impressive results in extracting visual representations from unlabeled imagery under a masking strategy. However, we reveal its disadvantages, notably its insufficient understanding of local semantics. This deficiency originates from masked modeling in the embedding space, resulting in a reduction of discriminative power and can even lead to the neglect of critical local semantics. To bridge this gap, we introduce DMT-JEPA, a novel masked modeling objective rooted in JEPA, specifically designed to generate discriminative latent targets from neighboring information. Our key idea is simple: we consider a set of semantically similar neighboring patches as a target of a masked patch. To be specific, the proposed DMT-JEPA (a) computes feature similarities between each masked patch and its corresponding neighboring patches to select patches having semantically meaningful relations, and (b) employs lightweight cross-attention heads to aggregate features of neighboring patches as the masked targets. Consequently, DMT-JEPA demonstrates strong discriminative power, offering benefits across a diverse spectrum of downstream tasks. Through extensive experiments, we demonstrate our effectiveness across various visual benchmarks, including ImageNet-1K image classification, ADE20K semantic segmentation, and COCO object detection tasks. Code is available at: \url{https://github.com/DMTJEPA/DMTJEPA}.
Scene Graph Generation (SGG) is a task that encodes visual relationships between objects in images as graph structures. SGG shows significant promise as a foundational component for downstream tasks, such as reasoning for embodied agents. To enable real-time applications, SGG must address the trade-off between performance and inference speed. However, current methods tend to focus on one of the following: (1) improving relation prediction accuracy, (2) enhancing object detection accuracy, or (3) reducing latency, without aiming to balance all three objectives simultaneously. To address this limitation, we propose a novel architecture, inference method, and relation prediction model. Our proposed solution, the REACT model, achieves the highest inference speed among existing SGG models, improving object detection accuracy without sacrificing relation prediction performance. Compared to state-of-the-art approaches, REACT is 2.7 times faster (with a latency of 23 ms) and improves object detection accuracy by 58.51%. Furthermore, our proposal significantly reduces model size, with an average of 5.5x fewer parameters. Code is available at https://github.com/Maelic/SGG-Benchmark
The automated surface defect detection is a fundamental task in industrial production, and the existing saliencybased works overcome the challenging scenes and give promising detection results. However, the cutting-edge efforts often suffer from large parameter size, heavy computational cost, and slow inference speed, which heavily limits the practical applications. To this end, we devise a multi-scale interactive (MI) module, which employs depthwise convolution (DWConv) and pointwise convolution (PWConv) to independently extract and interactively fuse features of different scales, respectively. Particularly, the MI module can provide satisfactory characterization for defect regions with fewer parameters. Embarking on this module, we propose a lightweight Multi-scale Interactive Network (MINet) to conduct real-time salient object detection of strip steel surface defects. Comprehensive experimental results on SD-Saliency-900 dataset, which contains three kinds of strip steel surface defect detection images (i.e., inclusion, patches, and scratches), demonstrate that the proposed MINet presents comparable detection accuracy with the state-of-the-art methods while running at a GPU speed of 721FPS and a CPU speed of 6.3FPS for 368*368 images with only 0.28M parameters. The code is available at https://github.com/Kunye-Shen/MINet.
Most recent multispectral object detectors employ a two-branch structure to extract features from RGB and thermal images. While the two-branch structure achieves better performance than a single-branch structure, it overlooks inference efficiency. This conflict is increasingly aggressive, as recent works solely pursue higher performance rather than both performance and efficiency. In this paper, we address this issue by improving the performance of efficient single-branch structures. We revisit the reasons causing the performance gap between these structures. For the first time, we reveal the information interference problem in the naive early-fusion strategy adopted by previous single-branch structures. Besides, we find that the domain gap between multispectral images, and weak feature representation of the single-branch structure are also key obstacles for performance. Focusing on these three problems, we propose corresponding solutions, including a novel shape-priority early-fusion strategy, a weakly supervised learning method, and a core knowledge distillation technique. Experiments demonstrate that single-branch networks equipped with these three contributions achieve significant performance enhancements while retaining high efficiency. Our code will be available at \url{https://github.com/XueZ-phd/Efficient-RGB-T-Early-Fusion-Detection}.
Ensuring robust 3D object detection and localization is crucial for many applications in robotics and autonomous driving. Recent models, however, face difficulties in maintaining high performance when applied to domains with differing sensor setups or geographic locations, often resulting in poor localization accuracy due to domain shift. To overcome this challenge, we introduce a novel diffusion-based box refinement approach. This method employs a domain-agnostic diffusion model, conditioned on the LiDAR points surrounding a coarse bounding box, to simultaneously refine the box's location, size, and orientation. We evaluate this approach under various domain adaptation settings, and our results reveal significant improvements across different datasets, object classes and detectors. Our PyTorch implementation is available at \href{https://github.com/cxy1997/DiffuBox}{https://github.com/cxy1997/DiffuBox}.
We present a novel synthetically generated multi-modal dataset, SCaRL, to enable the training and validation of autonomous driving solutions. Multi-modal datasets are essential to attain the robustness and high accuracy required by autonomous systems in applications such as autonomous driving. As deep learning-based solutions are becoming more prevalent for object detection, classification, and tracking tasks, there is great demand for datasets combining camera, lidar, and radar sensors. Existing real/synthetic datasets for autonomous driving lack synchronized data collection from a complete sensor suite. SCaRL provides synchronized Synthetic data from RGB, semantic/instance, and depth Cameras; Range-Doppler-Azimuth/Elevation maps and raw data from Radar; and 3D point clouds/2D maps of semantic, depth and Doppler data from coherent Lidar. SCaRL is a large dataset based on the CARLA Simulator, which provides data for diverse, dynamic scenarios and traffic conditions. SCaRL is the first dataset to include synthetic synchronized data from coherent Lidar and MIMO radar sensors. The dataset can be accessed here: https://fhr-ihs-sva.pages.fraunhofer.de/asp/scarl/