Deep Neural Networks (DNNs) are vulnerable to the black-box adversarial attack that is highly transferable. This threat comes from the distribution gap between adversarial and clean samples in feature space of the target DNNs. In this paper, we use Deep Generative Networks (DGNs) with a novel training mechanism to eliminate the distribution gap. The trained DGNs align the distribution of adversarial samples with clean ones for the target DNNs by translating pixel values. Different from previous work, we propose a more effective pixel level training constraint to make this achievable, thus enhancing robustness on adversarial samples. Further, a class-aware feature-level constraint is formulated for integrated distribution alignment. Our approach is general and applicable to multiple tasks, including image classification, semantic segmentation, and object detection. We conduct extensive experiments on different datasets. Our strategy demonstrates its unique effectiveness and generality against black-box attacks.
Depth cues are known to be useful for visual perception. However, direct measurement of depth is often impracticable. Fortunately, though, modern learning-based methods offer promising depth maps by inference in the wild. In this work, we adapt such depth inference models for object segmentation using the objects' "pop-out" prior in 3D. The "pop-out" is a simple composition prior that assumes objects reside on the background surface. Such compositional prior allows us to reason about objects in the 3D space. More specifically, we adapt the inferred depth maps such that objects can be localized using only 3D information. Such separation, however, requires knowledge about contact surface which we learn using the weak supervision of the segmentation mask. Our intermediate representation of contact surface, and thereby reasoning about objects purely in 3D, allows us to better transfer the depth knowledge into semantics. The proposed adaptation method uses only the depth model without needing the source data used for training, making the learning process efficient and practical. Our experiments on eight datasets of two challenging tasks, namely camouflaged object detection and salient object detection, consistently demonstrate the benefit of our method in terms of both performance and generalizability.
LiDAR and camera fusion techniques are promising for achieving 3D object detection in autonomous driving. Most multi-modal 3D object detection frameworks integrate semantic knowledge from 2D images into 3D LiDAR point clouds to enhance detection accuracy. Nevertheless, the restricted resolution of 2D feature maps impedes accurate re-projection and often induces a pronounced boundary-blurring effect, which is primarily attributed to erroneous semantic segmentation. To well handle this limitation, we propose a general multi-modal fusion framework Multi-Sem Fusion (MSF) to fuse the semantic information from both the 2D image and 3D points scene parsing results. Specifically, we employ 2D/3D semantic segmentation methods to generate the parsing results for 2D images and 3D point clouds. The 2D semantic information is further reprojected into the 3D point clouds with calibration parameters. To handle the misalignment between the 2D and 3D parsing results, we propose an Adaptive Attention-based Fusion (AAF) module to fuse them by learning an adaptive fusion score. Then the point cloud with the fused semantic label is sent to the following 3D object detectors. Furthermore, we propose a Deep Feature Fusion (DFF) module to aggregate deep features at different levels to boost the final detection performance. The effectiveness of the framework has been verified on two public large-scale 3D object detection benchmarks by comparing them with different baselines. The experimental results show that the proposed fusion strategies can significantly improve the detection performance compared to the methods using only point clouds and the methods using only 2D semantic information. Most importantly, the proposed approach significantly outperforms other approaches and sets state-of-the-art results on the nuScenes testing benchmark.
In object detection, post-processing methods like Non-maximum Suppression (NMS) are widely used. NMS can substantially reduce the number of false positive detections but may still keep some detections with low objectness scores. In order to find the exact number of objects and their labels in the image, we propose a post processing method called Detection Selection Algorithm (DSA) which is used after NMS or related methods. DSA greedily selects a subset of detected bounding boxes, together with full object reconstructions that give the interpretation of the whole image with highest likelihood, taking into account object occlusions. The algorithm consists of four components. First, we add an occlusion branch to Faster R-CNN to obtain occlusion relationships between objects. Second, we develop a single reconstruction algorithm which can reconstruct the whole appearance of an object given its visible part, based on the optimization of latent variables of a trained generative network which we call the decoder. Third, we propose a whole reconstruction algorithm which generates the joint reconstruction of all objects in a hypothesized interpretation, taking into account occlusion ordering. Finally we propose a greedy algorithm that incrementally adds or removes detections from a list to maximize the likelihood of the corresponding interpretation. DSA with NMS or Soft-NMS can achieve better results than NMS or Soft-NMS themselves, as is illustrated in our experiments on synthetic images with mutiple 3d objects.
Pedestrian detection in the wild remains a challenging problem especially when the scene contains significant occlusion and/or low resolution of the pedestrians to be detected. Existing methods are unable to adapt to these difficult cases while maintaining acceptable performance. In this paper we propose a novel feature learning model, referred to as CircleNet, to achieve feature adaptation by mimicking the process humans looking at low resolution and occluded objects: focusing on it again, at a finer scale, if the object can not be identified clearly for the first time. CircleNet is implemented as a set of feature pyramids and uses weight sharing path augmentation for better feature fusion. It targets at reciprocating feature adaptation and iterative object detection using multiple top-down and bottom-up pathways. To take full advantage of the feature adaptation capability in CircleNet, we design an instance decomposition training strategy to focus on detecting pedestrian instances of various resolutions and different occlusion levels in each cycle. Specifically, CircleNet implements feature ensemble with the idea of hard negative boosting in an end-to-end manner. Experiments on two pedestrian detection datasets, Caltech and CityPersons, show that CircleNet improves the performance of occluded and low-resolution pedestrians with significant margins while maintaining good performance on normal instances.
We present Recurrent Vision Transformers (RVTs), a novel backbone for object detection with event cameras. Event cameras provide visual information with sub-millisecond latency at a high-dynamic range and with strong robustness against motion blur. These unique properties offer great potential for low-latency object detection and tracking in time-critical scenarios. Prior work in event-based vision has achieved outstanding detection performance but at the cost of substantial inference time, typically beyond 40 milliseconds. By revisiting the high-level design of recurrent vision backbones, we reduce inference time by a factor of 6 while retaining similar performance. To achieve this, we explore a multi-stage design that utilizes three key concepts in each stage: First, a convolutional prior that can be regarded as a conditional positional embedding. Second, local and dilated global self-attention for spatial feature interaction. Third, recurrent temporal feature aggregation to minimize latency while retaining temporal information. RVTs can be trained from scratch to reach state-of-the-art performance on event-based object detection - achieving an mAP of 47.2% on the Gen1 automotive dataset. At the same time, RVTs offer fast inference (<12 ms on a T4 GPU) and favorable parameter efficiency (5 times fewer than prior art). Our study brings new insights into effective design choices that can be fruitful for research beyond event-based vision.
Detection Transformer (DETR) directly transforms queries to unique objects by using one-to-one bipartite matching during training and enables end-to-end object detection. Recently, these models have surpassed traditional detectors on COCO with undeniable elegance. However, they differ from traditional detectors in multiple designs, including model architecture and training schedules, and thus the effectiveness of one-to-one matching is not fully understood. In this work, we conduct a strict comparison between the one-to-one Hungarian matching in DETRs and the one-to-many label assignments in traditional detectors with non-maximum supervision (NMS). Surprisingly, we observe one-to-many assignments with NMS consistently outperform standard one-to-one matching under the same setting, with a significant gain of up to 2.5 mAP. Our detector that trains Deformable-DETR with traditional IoU-based label assignment achieved 50.2 COCO mAP within 12 epochs (1x schedule) with ResNet50 backbone, outperforming all existing traditional or transformer-based detectors in this setting. On multiple datasets, schedules, and architectures, we consistently show bipartite matching is unnecessary for performant detection transformers. Furthermore, we attribute the success of detection transformers to their expressive transformer architecture. Code is available at https://github.com/jozhang97/DETA.
Zero-shot detection (ZSD) is a challenging task where we aim to recognize and localize objects simultaneously, even when our model has not been trained with visual samples of a few target ("unseen") classes. Recently, methods employing generative models like GANs have shown some of the best results, where unseen-class samples are generated based on their semantics by a GAN trained on seen-class data, enabling vanilla object detectors to recognize unseen objects. However, the problem of semantic confusion still remains, where the model is sometimes unable to distinguish between semantically-similar classes. In this work, we propose to train a generative model incorporating a triplet loss that acknowledges the degree of dissimilarity between classes and reflects them in the generated samples. Moreover, a cyclic-consistency loss is also enforced to ensure that generated visual samples of a class highly correspond to their own semantics. Extensive experiments on two benchmark ZSD datasets - MSCOCO and PASCAL-VOC - demonstrate significant gains over the current ZSD methods, reducing semantic confusion and improving detection for the unseen classes.
Mitotic activity is key for the assessment of malignancy in many tumors. Moreover, it has been demonstrated that the proportion of abnormal mitosis to normal mitosis is of prognostic significance. Atypical mitotic figures (MF) can be identified morphologically as having segregation abnormalities of the chromatids. In this work, we perform, for the first time, automatic subtyping of mitotic figures into normal and atypical categories according to characteristic morphological appearances of the different phases of mitosis. Using the publicly available MIDOG21 and TUPAC16 breast cancer mitosis datasets, two experts blindly subtyped mitotic figures into five morphological categories. Further, we set up a state-of-the-art object detection pipeline extending the anchor-free FCOS approach with a gated hierarchical subclassification branch. Our labeling experiment indicated that subtyping of mitotic figures is a challenging task and prone to inter-rater disagreement, which we found in 24.89% of MF. Using the more diverse MIDOG21 dataset for training and TUPAC16 for testing, we reached a mean overall average precision score of 0.552, a ROC AUC score of 0.833 for atypical/normal MF and a mean class-averaged ROC-AUC score of 0.977 for discriminating the different phases of cells undergoing mitosis.
We propose a new self-supervised method for pre-training the backbone of deep perception models operating on point clouds. The core idea is to train the model on a pretext task which is the reconstruction of the surface on which the 3D points are sampled, and to use the underlying latent vectors as input to the perception head. The intuition is that if the network is able to reconstruct the scene surface, given only sparse input points, then it probably also captures some fragments of semantic information, that can be used to boost an actual perception task. This principle has a very simple formulation, which makes it both easy to implement and widely applicable to a large range of 3D sensors and deep networks performing semantic segmentation or object detection. In fact, it supports a single-stream pipeline, as opposed to most contrastive learning approaches, allowing training on limited resources. We conducted extensive experiments on various autonomous driving datasets, involving very different kinds of lidars, for both semantic segmentation and object detection. The results show the effectiveness of our method to learn useful representations without any annotation, compared to existing approaches. Code is available at https://github.com/valeoai/ALSO
The recent trend in 2D multiple object tracking (MOT) is jointly solving detection and tracking, where object detection and appearance feature (or motion) are learned simultaneously. Despite competitive performance, in crowded scenes, joint detection and tracking usually fail to find accurate object associations due to missed or false detections. In this paper, we jointly model counting, detection and re-identification in an end-to-end framework, named CountingMOT, tailored for crowded scenes. By imposing mutual object-count constraints between detection and counting, the CountingMOT tries to find a balance between object detection and crowd density map estimation, which can help it to recover missed detections or reject false detections. Our approach is an attempt to bridge the gap of object detection, counting, and re-Identification. This is in contrast to prior MOT methods that either ignore the crowd density and thus are prone to failure in crowded scenes,or depend on local correlations to build a graphical relationship for matching targets. The proposed MOT tracker can perform online and real-time tracking, and achieves the state-of-the-art results on public benchmarks MOT16 (MOTA of 79.7), MOT17 (MOTA of 81.3%) and MOT20 (MOTA of 78.9%).
The detection and prevention of illegal fishing is critical to maintaining a healthy and functional ecosystem. Recent research on ship detection in satellite imagery has focused exclusively on performance improvements, disregarding detection efficiency. However, the speed and compute cost of vessel detection are essential for a timely intervention to prevent illegal fishing. Therefore, we investigated optimization methods that lower detection time and cost with minimal performance loss. We trained an object detection model based on a convolutional neural network (CNN) using a dataset of satellite images. Then, we designed two efficiency optimizations that can be applied to the base CNN or any other base model. The optimizations consist of a fast, cheap classification model and a statistical algorithm. The integration of the optimizations with the object detection model leads to a trade-off between speed and performance. We studied the trade-off using metrics that give different weight to execution time and performance. We show that by using a classification model the average precision of the detection model can be approximated to 99.5% in 44% of the time or to 92.7% in 25% of the time.
Existing LiDAR-based 3D object detection methods for autonomous driving scenarios mainly adopt the training-from-scratch paradigm. Unfortunately, this paradigm heavily relies on large-scale labeled data, whose collection can be expensive and time-consuming. Self-supervised pre-training is an effective and desirable way to alleviate this dependence on extensive annotated data. In this work, we present BEV-MAE, an efficient masked autoencoder pre-training framework for LiDAR-based 3D object detection in autonomous driving. Specifically, we propose a bird's eye view (BEV) guided masking strategy to guide the 3D encoder learning feature representation in a BEV perspective and avoid complex decoder design during pre-training. Furthermore, we introduce a learnable point token to maintain a consistent receptive field size of the 3D encoder with fine-tuning for masked point cloud inputs. Based on the property of outdoor point clouds in autonomous driving scenarios, i.e., the point clouds of distant objects are more sparse, we propose point density prediction to enable the 3D encoder to learn location information, which is essential for object detection. Experimental results show that BEV-MAE surpasses prior state-of-the-art self-supervised methods and achieves a favorably pre-training efficiency. Furthermore, based on TransFusion-L, BEV-MAE achieves new state-of-the-art LiDAR-based 3D object detection results, with 73.6 NDS and 69.6 mAP on the nuScenes benchmark. The source code will be released at https://github.com/VDIGPKU/BEV-MAE
In light of the COVID-19 pandemic, patients were required to manually input their daily oxygen saturation (SpO2) and pulse rate (PR) values into a health monitoring system-unfortunately, such a process trend to be an error in typing. Several studies attempted to detect the physiological value from the captured image using optical character recognition (OCR). However, the technology has limited availability with high cost. Thus, this study aimed to propose a novel framework called PACMAN (Pandemic Accelerated Human-Machine Collaboration) with a low-resource deep learning-based computer vision. We compared state-of-the-art object detection algorithms (scaled YOLOv4, YOLOv5, and YOLOR), including the commercial OCR tools for digit recognition on the captured images from pulse oximeter display. All images were derived from crowdsourced data collection with varying quality and alignment. YOLOv5 was the best-performing model against the given model comparison across all datasets, notably the correctly orientated image dataset. We further improved the model performance with the digits auto-orientation algorithm and applied a clustering algorithm to extract SpO2 and PR values. The accuracy performance of YOLOv5 with the implementations was approximately 81.0-89.5%, which was enhanced compared to without any additional implementation. Accordingly, this study highlighted the completion of PACMAN framework to detect and read digits in real-world datasets. The proposed framework has been currently integrated into the patient monitoring system utilized by hospitals nationwide.
Fires have destructive power when they break out and affect their surroundings on a devastatingly large scale. The best way to minimize their damage is to detect the fire as quickly as possible before it has a chance to grow. Accordingly, this work looks into the potential of AI to detect and recognize fires and reduce detection time using object detection on an image stream. Object detection has made giant leaps in speed and accuracy over the last six years, making real-time detection feasible. To our end, we collected and labeled appropriate data from several public sources, which have been used to train and evaluate several models based on the popular YOLOv4 object detector. Our focus, driven by a collaborating industrial partner, is to implement our system in an industrial warehouse setting, which is characterized by high ceilings. A drawback of traditional smoke detectors in this setup is that the smoke has to rise to a sufficient height. The AI models brought forward in this research managed to outperform these detectors by a significant amount of time, providing precious anticipation that could help to minimize the effects of fires further.
LiDAR and camera are two essential sensors for 3D object detection in autonomous driving. LiDAR provides accurate and reliable 3D geometry information while the camera provides rich texture with color. Despite the increasing popularity of fusing these two complementary sensors, the challenge remains in how to effectively fuse 3D LiDAR point cloud with 2D camera images. Recent methods focus on point-level fusion which paints the LiDAR point cloud with camera features in the perspective view or bird's-eye view (BEV)-level fusion which unifies multi-modality features in the BEV representation. In this paper, we rethink these previous fusion strategies and analyze their information loss and influences on geometric and semantic features. We present SemanticBEVFusion to deeply fuse camera features with LiDAR features in a unified BEV representation while maintaining per-modality strengths for 3D object detection. Our method achieves state-of-the-art performance on the large-scale nuScenes dataset, especially for challenging distant objects. The code will be made publicly available.
Contrastive learning has emerged as a competitive pretraining method for object detection. Despite this progress, there has been minimal investigation into the robustness of contrastively pretrained detectors when faced with domain shifts. To address this gap, we conduct an empirical study of contrastive learning and out-of-domain object detection, studying how contrastive view design affects robustness. In particular, we perform a case study of the detection-focused pretext task Instance Localization (InsLoc) and propose strategies to augment views and enhance robustness in appearance-shifted and context-shifted scenarios. Amongst these strategies, we propose changes to cropping such as altering the percentage used, adding IoU constraints, and integrating saliency based object priors. We also explore the addition of shortcut-reducing augmentations such as Poisson blending, texture flattening, and elastic deformation. We benchmark these strategies on abstract, weather, and context domain shifts and illustrate robust ways to combine them, in both pretraining on single-object and multi-object image datasets. Overall, our results and insights show how to ensure robustness through the choice of views in contrastive learning.
Most camera lens systems are designed in isolation, separately from downstream computer vision methods. Recently, joint optimization approaches that design lenses alongside other components of the image acquisition and processing pipeline -- notably, downstream neural networks -- have achieved improved imaging quality or better performance on vision tasks. However, these existing methods optimize only a subset of lens parameters and cannot optimize glass materials given their categorical nature. In this work, we develop a differentiable spherical lens simulation model that accurately captures geometrical aberrations. We propose an optimization strategy to address the challenges of lens design -- notorious for non-convex loss function landscapes and many manufacturing constraints -- that are exacerbated in joint optimization tasks. Specifically, we introduce quantized continuous glass variables to facilitate the optimization and selection of glass materials in an end-to-end design context, and couple this with carefully designed constraints to support manufacturability. In automotive object detection, we report improved detection performance over existing designs even when simplifying designs to two- or three-element lenses, despite significantly degrading the image quality.
In this paper, we introduce a novel approach for ground plane normal estimation of wheeled vehicles. In practice, the ground plane is dynamically changed due to braking and unstable road surface. As a result, the vehicle pose, especially the pitch angle, is oscillating from subtle to obvious. Thus, estimating ground plane normal is meaningful since it can be encoded to improve the robustness of various autonomous driving tasks (e.g., 3D object detection, road surface reconstruction, and trajectory planning). Our proposed method only uses odometry as input and estimates accurate ground plane normal vectors in real time. Particularly, it fully utilizes the underlying connection between the ego pose odometry (ego-motion) and its nearby ground plane. Built on that, an Invariant Extended Kalman Filter (IEKF) is designed to estimate the normal vector in the sensor's coordinate. Thus, our proposed method is simple yet efficient and supports both camera- and inertial-based odometry algorithms. Its usability and the marked improvement of robustness are validated through multiple experiments on public datasets. For instance, we achieve state-of-the-art accuracy on KITTI dataset with the estimated vector error of 0.39{\deg}. Our code is available at github.com/manymuch/ground_normal_filter.
Object detection requires substantial labeling effort for learning robust models. Active learning can reduce this effort by intelligently selecting relevant examples to be annotated. However, selecting these examples properly without introducing a sampling bias with a negative impact on the generalization performance is not straightforward and most active learning techniques can not hold their promises on real-world benchmarks. In our evaluation paper, we focus on active learning techniques without a computational overhead besides inference, something we refer to as zero-cost active learning. In particular, we show that a key ingredient is not only the score on a bounding box level but also the technique used for aggregating the scores for ranking images. We outline our experimental setup and also discuss practical considerations when using active learning for object detection.