There is a surge of interest in image scene graph generation (object, attribute and relationship detection) due to the need of building fine-grained image understanding models that go beyond object detection. Due to the lack of a good benchmark, the reported results of different scene graph generation models are not directly comparable, impeding the research progress. We have developed a much-needed scene graph generation benchmark based on the maskrcnn-benchmark and several popular models. This paper presents main features of our benchmark and a comprehensive ablation study of scene graph generation models using the Visual Genome and OpenImages Visual relationship detection datasets. Our codebase is made publicly available at https://github.com/microsoft/scene_graph_benchmark.
We propose Rank & Sort (RS) Loss, a ranking-based loss function to train deep object detection and instance segmentation methods (i.e. visual detectors). RS Loss supervises the classifier, a sub-network of these methods, to rank each positive above all negatives as well as to sort positives among themselves with respect to (wrt.) their localisation qualities (e.g. Intersection-over-Union - IoU). To tackle the non-differentiable nature of ranking and sorting, we reformulate the incorporation of error-driven update with backpropagation as Identity Update, which enables us to model our novel sorting error among positives. With RS Loss, we significantly simplify training: (i) Thanks to our sorting objective, the positives are prioritized by the classifier without an additional auxiliary head (e.g. for centerness, IoU, mask-IoU), (ii) due to its ranking-based nature, RS Loss is robust to class imbalance, and thus, no sampling heuristic is required, and (iii) we address the multi-task nature of visual detectors using tuning-free task-balancing coefficients. Using RS Loss, we train seven diverse visual detectors only by tuning the learning rate, and show that it consistently outperforms baselines: e.g. our RS Loss improves (i) Faster R-CNN by ~ 3 box AP and aLRP Loss (ranking-based baseline) by ~ 2 box AP on COCO dataset, (ii) Mask R-CNN with repeat factor sampling (RFS) by 3.5 mask AP (~ 7 AP for rare classes) on LVIS dataset; and also outperforms all counterparts. Code is available at: https://github.com/kemaloksuz/RankSortLoss
Exploring to what humans pay attention in dynamic panoramic scenes is useful for many fundamental applications, including augmented reality (AR) in retail, AR-powered recruitment, and visual language navigation. With this goal in mind, we propose PV-SOD, a new task that aims to segment salient objects from panoramic videos. In contrast to existing fixation-/object-level saliency detection tasks, we focus on audio-induced salient object detection (SOD), where the salient objects are labeled with the guidance of audio-induced eye movements. To support this task, we collect the first large-scale dataset, named ASOD60K, which contains 4K-resolution video frames annotated with a six-level hierarchy, thus distinguishing itself with richness, diversity and quality. Specifically, each sequence is marked with both its super-/sub-class, with objects of each sub-class being further annotated with human eye fixations, bounding boxes, object-/instance-level masks, and associated attributes (e.g., geometrical distortion). These coarse-to-fine annotations enable detailed analysis for PV-SOD modelling, e.g., determining the major challenges for existing SOD models, and predicting scanpaths to study the long-term eye fixation behaviors of humans. We systematically benchmark 11 representative approaches on ASOD60K and derive several interesting findings. We hope this study could serve as a good starting point for advancing SOD research towards panoramic videos. The dataset and benchmark will be made publicly available at https://github.com/PanoAsh/ASOD60K.
Transformer with self-attention has led to the revolutionizing of natural language processing field, and recently inspires the emergence of Transformer-style architecture design with competitive results in numerous computer vision tasks. Nevertheless, most of existing designs directly employ self-attention over a 2D feature map to obtain the attention matrix based on pairs of isolated queries and keys at each spatial location, but leave the rich contexts among neighbor keys under-exploited. In this work, we design a novel Transformer-style module, i.e., Contextual Transformer (CoT) block, for visual recognition. Such design fully capitalizes on the contextual information among input keys to guide the learning of dynamic attention matrix and thus strengthens the capacity of visual representation. Technically, CoT block first contextually encodes input keys via a $3\times3$ convolution, leading to a static contextual representation of inputs. We further concatenate the encoded keys with input queries to learn the dynamic multi-head attention matrix through two consecutive $1\times1$ convolutions. The learnt attention matrix is multiplied by input values to achieve the dynamic contextual representation of inputs. The fusion of the static and dynamic contextual representations are finally taken as outputs. Our CoT block is appealing in the view that it can readily replace each $3\times3$ convolution in ResNet architectures, yielding a Transformer-style backbone named as Contextual Transformer Networks (CoTNet). Through extensive experiments over a wide range of applications (e.g., image recognition, object detection and instance segmentation), we validate the superiority of CoTNet as a stronger backbone. Source code is available at \url{https://github.com/JDAI-CV/CoTNet}.
Transparent objects are a very challenging problem in computer vision. They are hard to segment or classify due to their lack of precise boundaries, and there is limited data available for training deep neural networks. As such, current solutions for this problem employ rigid synthetic datasets, which lack flexibility and lead to severe performance degradation when deployed on real-world scenarios. In particular, these synthetic datasets omit features such as refraction, dispersion and caustics due to limitations in the rendering pipeline. To address this issue, we present SuperCaustics, a real-time, open-source simulation of transparent objects designed for deep learning applications. SuperCaustics features extensive modules for stochastic environment creation; uses hardware ray-tracing to support caustics, dispersion, and refraction; and enables generating massive datasets with multi-modal, pixel-perfect ground truth annotations. To validate our proposed system, we trained a deep neural network from scratch to segment transparent objects in difficult lighting scenarios. Our neural network achieved performance comparable to the state-of-the-art on a real-world dataset using only 10% of the training data and in a fraction of the training time. Further experiments show that a model trained with SuperCaustics can segment different types of caustics, even in images with multiple overlapping transparent objects. To the best of our knowledge, this is the first such result for a model trained on synthetic data. Both our open-source code and experimental data are freely available online.
This paper presents a simple MLP-like architecture, CycleMLP, which is a versatile backbone for visual recognition and dense predictions. As compared to modern MLP architectures, e.g., MLP-Mixer, ResMLP, and gMLP, whose architectures are correlated to image size and thus are infeasible in object detection and segmentation, CycleMLP has two advantages compared to modern approaches. (1) It can cope with various image sizes. (2) It achieves linear computational complexity to image size by using local windows. In contrast, previous MLPs have $O(N^2)$ computations due to fully spatial connections. We build a family of models which surpass existing MLPs and even state-of-the-art Transformer-based models, e.g., Swin Transformer, while using fewer parameters and FLOPs. We expand the MLP-like models' applicability, making them a versatile backbone for dense prediction tasks. CycleMLP achieves competitive results on object detection, instance segmentation, and semantic segmentation. In particular, CycleMLP-Tiny outperforms Swin-Tiny by 1.3% mIoU on ADE20K dataset with fewer FLOPs. Moreover, CycleMLP also shows excellent zero-shot robustness on ImageNet-C dataset. Code is available at https://github.com/ShoufaChen/CycleMLP.
In recent years, many semantic segmentation methods have been proposed to predict label of pixels in the scene. In general, we measure area prediction errors or boundary prediction errors for comparing methods. However, there is no intuitive evaluation metric that evaluates both aspects. In this work, we propose a new evaluation measure called weighted Intersection over Union (wIoU) for semantic segmentation. First, it builds a weight map generated from a boundary distance map, allowing weighted evaluation for each pixel based on a boundary importance factor. The proposed wIoU can evaluate both contour and region by setting a boundary importance factor. We validated the effectiveness of wIoU on a dataset of 33 scenes and demonstrated its flexibility. Using the proposed metric, we expect more flexible and intuitive evaluation in semantic segmentation field are possible.
Meningioma is one of the most prevalent brain tumors in adults. To determine its malignancy, it is graded by a pathologist into three grades according to WHO standards. This grade plays a decisive role in treatment, and yet may be subject to inter-rater discordance. In this work, we present and compare three approaches towards fully automatic meningioma grading from histology whole slide images. All approaches are following a two-stage paradigm, where we first identify a region of interest based on the detection of mitotic figures in the slide using a state-of-the-art object detection deep learning network. This region of highest mitotic rate is considered characteristic for biological tumor behavior. In the second stage, we calculate a score corresponding to tumor malignancy based on information contained in this region using three different settings. In a first approach, image patches are sampled from this region and regression is based on morphological features encoded by a ResNet-based network. We compare this to learning a logistic regression from the determined mitotic count, an approach which is easily traceable and explainable. Lastly, we combine both approaches in a single network. We trained the pipeline on 951 slides from 341 patients and evaluated them on a separate set of 141 slides from 43 patients. All approaches yield a high correlation to the WHO grade. The logistic regression and the combined approach had the best results in our experiments, yielding correct predictions in 32 and 33 of all cases, respectively, with the image-based approach only predicting 25 cases correctly. Spearman's correlation was 0.716, 0.792 and 0.790 respectively. It may seem counterintuitive at first that morphological features provided by image patches do not improve model performance. Yet, this mirrors the criteria of the grading scheme, where mitotic count is the only unequivocal parameter.
In this report, we present some experienced improvements to YOLO series, forming a new high-performance detector -- YOLOX. We switch the YOLO detector to an anchor-free manner and conduct other advanced detection techniques, i.e., a decoupled head and the leading label assignment strategy SimOTA to achieve state-of-the-art results across a large scale range of models: For YOLO-Nano with only 0.91M parameters and 1.08G FLOPs, we get 25.3% AP on COCO, surpassing NanoDet by 1.8% AP; for YOLOv3, one of the most widely used detectors in industry, we boost it to 47.3% AP on COCO, outperforming the current best practice by 3.0% AP; for YOLOX-L with roughly the same amount of parameters as YOLOv4-CSP, YOLOv5-L, we achieve 50.0% AP on COCO at a speed of 68.9 FPS on Tesla V100, exceeding YOLOv5-L by 1.8% AP. Further, we won the 1st Place on Streaming Perception Challenge (Workshop on Autonomous Driving at CVPR 2021) using a single YOLOX-L model. We hope this report can provide useful experience for developers and researchers in practical scenes, and we also provide deploy versions with ONNX, TensorRT, NCNN, and Openvino supported. Source code is at https://github.com/Megvii-BaseDetection/YOLOX.
An Axial Shifted MLP architecture (AS-MLP) is proposed in this paper. Different from MLP-Mixer, where the global spatial feature is encoded for information flow through matrix transposition and one token-mixing MLP, we pay more attention to the local features interaction. By axially shifting channels of the feature map, AS-MLP is able to obtain the information flow from different axial directions, which captures the local dependencies. Such an operation enables us to utilize a pure MLP architecture to achieve the same local receptive field as CNN-like architecture. We can also design the receptive field size and dilation of blocks of AS-MLP, etc, in the same spirit of convolutional neural networks. With the proposed AS-MLP architecture, our model obtains 83.3% Top-1 accuracy with 88M parameters and 15.2 GFLOPs on the ImageNet-1K dataset. Such a simple yet effective architecture outperforms all MLP-based architectures and achieves competitive performance compared to the transformer-based architectures (e.g., Swin Transformer) even with slightly lower FLOPs. In addition, AS-MLP is also the first MLP-based architecture to be applied to the downstream tasks (e.g., object detection and semantic segmentation). The experimental results are also impressive. Our proposed AS-MLP obtains 51.5 mAP on the COCO validation set and 49.5 MS mIoU on the ADE20K dataset, which is competitive compared to the transformer-based architectures. Our AS-MLP establishes a strong baseline of MLP-based architecture. Code is available at https://github.com/svip-lab/AS-MLP.
Lidar-based object detectors are critical parts of the 3D perception pipeline in autonomous navigation systems such as self-driving cars. However, they are known to be sensitive to adverse weather conditions such as rain, snow and fog due to reduced signal-to-noise ratio (SNR) and signal-to-background ratio (SBR). As a result, lidar-based object detectors trained on data captured in normal weather tend to perform poorly in such scenarios. However, collecting and labelling sufficient training data in a diverse range of adverse weather conditions is laborious and prohibitively expensive. To address this issue, we propose a physics-based approach to simulate lidar point clouds of scenes in adverse weather conditions. These augmented datasets can then be used to train lidar-based detectors to improve their all-weather reliability. Specifically, we introduce a hybrid Monte-Carlo based approach that treats (i) the effects of large particles by placing them randomly and comparing their back reflected power against the target, and (ii) attenuation effects on average through calculation of scattering efficiencies from the Mie theory and particle size distributions. Retraining networks with this augmented data improves mean average precision evaluated on real world rainy scenes and we observe greater improvement in performance with our model relative to existing models from the literature. Furthermore, we evaluate recent state-of-the-art detectors on the simulated weather conditions and present an in-depth analysis of their performance.
Human vision is able to capture the part-whole hierarchical information from the entire scene. This paper presents the Visual Parser (ViP) that explicitly constructs such a hierarchy with transformers. ViP divides visual representations into two levels, the part level and the whole level. Information of each part represents a combination of several independent vectors within the whole. To model the representations of the two levels, we first encode the information from the whole into part vectors through an attention mechanism, then decode the global information within the part vectors back into the whole representation. By iteratively parsing the two levels with the proposed encoder-decoder interaction, the model can gradually refine the features on both levels. Experimental results demonstrate that ViP can achieve very competitive performance on three major tasks e.g. classification, detection and instance segmentation. In particular, it can surpass the previous state-of-the-art CNN backbones by a large margin on object detection. The tiny model of the ViP family with $7.2\times$ fewer parameters and $10.9\times$ fewer FLOPS can perform comparably with the largest model ResNeXt-101-64$\times$4d of ResNe(X)t family. Visualization results also demonstrate that the learnt parts are highly informative of the predicting class, making ViP more explainable than previous fundamental architectures. Code is available at https://github.com/kevin-ssy/ViP.
Self-Attention has become prevalent in computer vision models. Inspired by fully connected Conditional Random Fields (CRFs), we decompose self-attention into local and context terms. They correspond to the unary and binary terms in CRF and are implemented by attention mechanisms with projection matrices. We observe that the unary terms only make small contributions to the outputs, and meanwhile standard CNNs that rely solely on the unary terms achieve great performances on a variety of tasks. Therefore, we propose Locally Enhanced Self-Attention (LESA), which enhances the unary term by incorporating it with convolutions, and utilizes a fusion module to dynamically couple the unary and binary operations. In our experiments, we replace the self-attention modules with LESA. The results on ImageNet and COCO show the superiority of LESA over convolution and self-attention baselines for the tasks of image recognition, object detection, and instance segmentation. The code is made publicly available.
Due to the vulnerability of deep neural networks to adversarial examples, numerous works on adversarial attacks and defenses have been burgeoning over the past several years. However, there seem to be some conventional views regarding adversarial attacks and object detection approaches that most researchers take for granted. In this work, we bring a fresh perspective on those procedures by evaluating the impact of universal perturbations on object detection at a class-level. We apply it to a carefully curated data set related to autonomous driving. We use Faster-RCNN object detector on images of five different categories: person, car, truck, stop sign and traffic light from the COCO data set, while carefully perturbing the images using Universal Dense Object Suppression algorithm. Our results indicate that person, car, traffic light, truck and stop sign are resilient in that order (most to least) to universal perturbations. To the best of our knowledge, this is the first time such a ranking has been established which is significant for the security of the data sets pertaining to autonomous vehicles and object detection in general.
The proliferation of remote sensing satellites has resulted in a massive amount of remote sensing images. However, due to human and material resource constraints, the vast majority of remote sensing images remain unlabeled. As a result, it cannot be applied to currently available deep learning methods. To fully utilize the remaining unlabeled images, we propose a Geographical Knowledge-driven Representation learning method for remote sensing images (GeoKR), improving network performance and reduce the demand for annotated data. The global land cover products and geographical location associated with each remote sensing image are regarded as geographical knowledge to provide supervision for representation learning and network pre-training. An efficient pre-training framework is proposed to eliminate the supervision noises caused by imaging times and resolutions difference between remote sensing images and geographical knowledge. A large scale pre-training dataset Levir-KR is proposed to support network pre-training. It contains 1,431,950 remote sensing images from Gaofen series satellites with various resolutions. Experimental results demonstrate that our proposed method outperforms ImageNet pre-training and self-supervised representation learning methods and significantly reduces the burden of data annotation on downstream tasks such as scene classification, semantic segmentation, object detection, and cloud / snow detection. It demonstrates that our proposed method can be used as a novel paradigm for pre-training neural networks. Codes will be available on https://github.com/flyakon/Geographical-Knowledge-driven-Representaion-Learning.
The development of lightweight object detectors is essential due to the limited computation resources. To reduce the computation cost, how to generate redundant features plays a significant role. This paper proposes a new lightweight Convolution method Cross-Stage Lightweight (CSL) Module, to generate redundant features from cheap operations. In the intermediate expansion stage, we replaced Pointwise Convolution with Depthwise Convolution to produce candidate features. The proposed CSL-Module can reduce the computation cost significantly. Experiments conducted at MS-COCO show that the proposed CSL-Module can approximate the fitting ability of Convolution-3x3. Finally, we use the module to construct a lightweight detector CSL-YOLO, achieving better detection performance with only 43% FLOPs and 52% parameters than Tiny-YOLOv4.
Object detection in autonomous cars is commonly based on camera images and Lidar inputs, which are often used to train prediction models such as deep artificial neural networks for decision making for object recognition, adjusting speed, etc. A mistake in such decision making can be damaging; thus, it is vital to measure the reliability of decisions made by such prediction models via uncertainty measurement. Uncertainty, in deep learning models, is often measured for classification problems. However, deep learning models in autonomous driving are often multi-output regression models. Hence, we propose a novel method called PURE (Prediction sURface uncErtainty) for measuring prediction uncertainty of such regression models. We formulate the object recognition problem as a regression model with more than one outputs for finding object locations in a 2-dimensional camera view. For evaluation, we modified three widely-applied object recognition models (i.e., YoLo, SSD300 and SSD512) and used the KITTI, Stanford Cars, Berkeley DeepDrive, and NEXET datasets. Results showed the statistically significant negative correlation between prediction surface uncertainty and prediction accuracy suggesting that uncertainty significantly impacts the decisions made by autonomous driving.
The vast majority of uncertainty quantification methods for deep object detectors such as variational inference are based on the network output. Here, we study gradient-based epistemic uncertainty metrics for deep object detectors to obtain reliable confidence estimates. We show that they contain predictive information and that they capture information orthogonal to that of common, output-based uncertainty estimation methods like Monte-Carlo dropout and deep ensembles. To this end, we use meta classification and meta regression to produce confidence estimates using gradient metrics and other baselines for uncertainty quantification which are in principle applicable to any object detection architecture. Specifically, we employ false positive detection and prediction of localization quality to investigate uncertainty content of our metrics and compute the calibration errors of meta classifiers. Moreover, we use them as a post-processing filter mechanism to the object detection pipeline and compare object detection performance. Our results show that gradient-based uncertainty is itself on par with output-based methods across different detectors and datasets. More significantly, combined meta classifiers based on gradient and output-based metrics outperform the standalone models. Based on this result, we conclude that gradient uncertainty adds orthogonal information to output-based methods. This suggests that variational inference may be supplemented by gradient-based uncertainty to obtain improved confidence measures, contributing to down-stream applications of deep object detectors and improving their probabilistic reliability.
Point clouds and RGB images are naturally complementary modalities for 3D visual understanding - the former provides sparse but accurate locations of points on objects, while the latter contains dense color and texture information. Despite this potential for close sensor fusion, many methods train two models in isolation and use simple feature concatenation to represent 3D sensor data. This separated training scheme results in potentially sub-optimal performance and prevents 3D tasks from being used to benefit 2D tasks that are often useful on their own. To provide a more integrated approach, we propose a novel Multi-Modality Task Cascade network (MTC-RCNN) that leverages 3D box proposals to improve 2D segmentation predictions, which are then used to further refine the 3D boxes. We show that including a 2D network between two stages of 3D modules significantly improves both 2D and 3D task performance. Moreover, to prevent the 3D module from over-relying on the overfitted 2D predictions, we propose a dual-head 2D segmentation training and inference scheme, allowing the 2nd 3D module to learn to interpret imperfect 2D segmentation predictions. Evaluating our model on the challenging SUN RGB-D dataset, we improve upon state-of-the-art results of both single modality and fusion networks by a large margin ($\textbf{+3.8}$ [email omitted]). Code will be released $\href{https://github.com/Divadi/MTC_RCNN}{\text{here.}}$
The detection of semantic relationships between objects represented in an image is one of the fundamental challenges in image interpretation. Neural-Symbolic techniques, such as Logic Tensor Networks (LTNs), allow the combination of semantic knowledge representation and reasoning with the ability to efficiently learn from examples typical of neural networks. We here propose Faster-LTN, an object detector composed of a convolutional backbone and an LTN. To the best of our knowledge, this is the first attempt to combine both frameworks in an end-to-end training setting. This architecture is trained by optimizing a grounded theory which combines labelled examples with prior knowledge, in the form of logical axioms. Experimental comparisons show competitive performance with respect to the traditional Faster R-CNN architecture.