We present DistillFlow, a knowledge distillation approach to learning optical flow. DistillFlow trains multiple teacher models and a student model, where challenging transformations are applied to the input of the student model to generate hallucinated occlusions as well as less confident predictions. Then, a self-supervised learning framework is constructed: confident predictions from teacher models are served as annotations to guide the student model to learn optical flow for those less confident predictions. The self-supervised learning framework enables us to effectively learn optical flow from unlabeled data, not only for non-occluded pixels, but also for occluded pixels. DistillFlow achieves state-of-the-art unsupervised learning performance on both KITTI and Sintel datasets. Our self-supervised pre-trained model also provides an excellent initialization for supervised fine-tuning, suggesting an alternate training paradigm in contrast to current supervised learning methods that highly rely on pre-training on synthetic data. At the time of writing, our fine-tuned models ranked 1st among all monocular methods on the KITTI 2015 benchmark, and outperform all published methods on the Sintel Final benchmark. More importantly, we demonstrate the generalization capability of DistillFlow in three aspects: framework generalization, correspondence generalization and cross-dataset generalization.
The field of neuromorphic computing promises extremely low-power and low-latency sensing and processing. Challenges in transferring learning algorithms from traditional artificial neural networks (ANNs) to spiking neural networks (SNNs) have so far prevented their application to large-scale, complex regression tasks. Furthermore, realizing a truly asynchronous and fully neuromorphic pipeline that maximally attains the abovementioned benefits involves rethinking the way in which this pipeline takes in and accumulates information. In the case of perception, spikes would be passed as-is and one-by-one between an event camera and an SNN, meaning all temporal integration of information must happen inside the network. In this article, we tackle these two problems. We focus on the complex task of learning to estimate optical flow from event-based camera inputs in a self-supervised manner, and modify the state-of-the-art ANN training pipeline to encode minimal temporal information in its inputs. Moreover, we reformulate the self-supervised loss function for event-based optical flow to improve its convexity. We perform experiments with various types of recurrent ANNs and SNNs using the proposed pipeline. Concerning SNNs, we investigate the effects of elements such as parameter initialization and optimization, surrogate gradient shape, and adaptive neuronal mechanisms. We find that initialization and surrogate gradient width play a crucial part in enabling learning with sparse inputs, while the inclusion of adaptivity and learnable neuronal parameters can improve performance. We show that the performance of the proposed ANNs and SNNs are on par with that of the current state-of-the-art ANNs trained in a self-supervised manner.
Estimating geometric elements such as depth, camera motion, and optical flow from images is an important part of the robot's visual perception. We use a joint self-supervised method to estimate the three geometric elements. Depth network, optical flow network and camera motion network are independent of each other but are jointly optimized during training phase. Compared with independent training, joint training can make full use of the geometric relationship between geometric elements and provide dynamic and static information of the scene. In this paper, we improve the joint self-supervision method from three aspects: network structure, dynamic object segmentation, and geometric constraints. In terms of network structure, we apply the attention mechanism to the camera motion network, which helps to take advantage of the similarity of camera movement between frames. And according to attention mechanism in Transformer, we propose a plug-and-play convolutional attention module. In terms of dynamic object, according to the different influences of dynamic objects in the optical flow self-supervised framework and the depth-pose self-supervised framework, we propose a threshold algorithm to detect dynamic regions, and mask that in the loss function respectively. In terms of geometric constraints, we use traditional methods to estimate the fundamental matrix from the corresponding points to constrain the camera motion network. We demonstrate the effectiveness of our method on the KITTI dataset. Compared with other joint self-supervised methods, our method achieves state-of-the-art performance in the estimation of pose and optical flow, and the depth estimation has also achieved competitive results. Code will be available https://github.com/jianfenglihg/Unsupervised_geometry.
In recent years, the LiDAR images, as a 2D compact representation of 3D LiDAR point clouds, are widely applied in various tasks, e.g., 3D semantic segmentation, LiDAR point cloud compression (PCC). Among these works, the optical flow estimation for LiDAR image sequences has become a key issue, especially for the motion estimation of the inter prediction in PCC. However, the existing optical flow estimation models are likely to be unreliable for LiDAR images. In this work, we first propose a light-weight flow estimation model for LiDAR image sequences. The key novelty of our method lies in two aspects. One is that for the different characteristics (with the spatial-variation feature distribution) of the LiDAR images w.r.t. the normal color images, we introduce the attention mechanism into our model to improve the quality of the estimated flow. The other one is that to tackle the lack of large-scale LiDAR-image annotations, we present an unsupervised method, which directly minimizes the inconsistency between the reference image and the reconstructed image based on the estimated optical flow. Extensive experimental results have shown that our proposed model outperforms other mainstream models on the KITTI dataset, with much fewer parameters.
We present a Python-based renderer built on NVIDIA's OptiX ray tracing engine and the OptiX AI denoiser, designed to generate high-quality synthetic images for research in computer vision and deep learning. Our tool enables the description and manipulation of complex dynamic 3D scenes containing object meshes, materials, textures, lighting, volumetric data (e.g., smoke), and backgrounds. Metadata, such as 2D/3D bounding boxes, segmentation masks, depth maps, normal maps, material properties, and optical flow vectors, can also be generated. In this work, we discuss design goals, architecture, and performance. We demonstrate the use of data generated by path tracing for training an object detector and pose estimator, showing improved performance in sim-to-real transfer in situations that are difficult for traditional raster-based renderers. We offer this tool as an easy-to-use, performant, high-quality renderer for advancing research in synthetic data generation and deep learning.
4D seismic imaging has been widely used in CO$_2$ sequestration projects to monitor the fluid flow in the volumetric subsurface region that is not sampled by wells. Ideally, real-time monitoring and near-future forecasting would provide site operators with great insights to understand the dynamics of the subsurface reservoir and assess any potential risks. However, due to obstacles such as high deployment cost, availability of acquisition equipment, exclusion zones around surface structures, only very sparse seismic imaging data can be obtained during monitoring. That leads to an unavoidable and growing knowledge gap over time. The operator needs to understand the fluid flow throughout the project lifetime and the seismic data are only available at a limited number of times. This is insufficient for understanding the reservoir behavior. To overcome those challenges, we have developed spatio-temporal neural-network-based models that can produce high-fidelity interpolated or extrapolated images effectively and efficiently. Specifically, our models are built on an autoencoder, and incorporate the long short-term memory (LSTM) structure with a new loss function regularized by optical flow. We validate the performance of our models using real 4D post-stack seismic imaging data acquired at the Sleipner CO$_2$ sequestration field. We employ two different strategies in evaluating our models. Numerically, we compare our models with different baseline approaches using classic pixel-based metrics. We also conduct a blind survey and collect a total of 20 responses from domain experts to evaluate the quality of data generated by our models. Via both numerical and expert evaluation, we conclude that our models can produce high-quality 2D/3D seismic imaging data at a reasonable cost, offering the possibility of real-time monitoring or even near-future forecasting of the CO$_2$ storage reservoir.
Recently, DETR and Deformable DETR have been proposed to eliminate the need for many hand-designed components in object detection while demonstrating good performance as previous complex hand-crafted detectors. However, their performance on Video Object Detection (VOD) has not been well explored. In this paper, we present TransVOD, an end-to-end video object detection model based on a spatial-temporal Transformer architecture. The goal of this paper is to streamline the pipeline of VOD, effectively removing the need for many hand-crafted components for feature aggregation, e.g., optical flow, recurrent neural networks, relation networks. Besides, benefited from the object query design in DETR, our method does not need complicated post-processing methods such as Seq-NMS or Tubelet rescoring, which keeps the pipeline simple and clean. In particular, we present temporal Transformer to aggregate both the spatial object queries and the feature memories of each frame. Our temporal Transformer consists of three components: Temporal Deformable Transformer Encoder (TDTE) to encode the multiple frame spatial details, Temporal Query Encoder (TQE) to fuse object queries, and Temporal Deformable Transformer Decoder to obtain current frame detection results. These designs boost the strong baseline deformable DETR by a significant margin (3%-4% mAP) on the ImageNet VID dataset. TransVOD yields comparable results performance on the benchmark of ImageNet VID. We hope our TransVOD can provide a new perspective for video object detection. Code will be made publicly available at https://github.com/SJTU-LuHe/TransVOD.
The technology for Visual Odometry (VO) that estimates the position and orientation of the moving object through analyzing the image sequences captured by on-board cameras, has been well investigated with the rising interest in autonomous driving. This paper studies monocular VO from the perspective of Deep Learning (DL). Unlike most current learning-based methods, our approach, called DeepAVO, is established on the intuition that features contribute discriminately to different motion patterns. Specifically, we present a novel four-branch network to learn the rotation and translation by leveraging Convolutional Neural Networks (CNNs) to focus on different quadrants of optical flow input. To enhance the ability of feature selection, we further introduce an effective channel-spatial attention mechanism to force each branch to explicitly distill related information for specific Frame to Frame (F2F) motion estimation. Experiments on various datasets involving outdoor driving and indoor walking scenarios show that the proposed DeepAVO outperforms the state-of-the-art monocular methods by a large margin, demonstrating competitive performance to the stereo VO algorithm and verifying promising potential for generalization.
Fingerspelling in sign language has been the means of communicating technical terms and proper nouns when they do not have dedicated sign language gestures. Automatic recognition of fingerspelling can help resolve communication barriers when interacting with deaf people. The main challenges prevalent in fingerspelling recognition are the ambiguity in the gestures and strong articulation of the hands. The automatic recognition model should address high inter-class visual similarity and high intra-class variation in the gestures. Most of the existing research in fingerspelling recognition has focused on the dataset collected in a controlled environment. The recent collection of a large-scale annotated fingerspelling dataset in the wild, from social media and online platforms, captures the challenges in a real-world scenario. In this work, we propose a fine-grained visual attention mechanism using the Transformer model for the sequence-to-sequence prediction task in the wild dataset. The fine-grained attention is achieved by utilizing the change in motion of the video frames (optical flow) in sequential context-based attention along with a Transformer encoder model. The unsegmented continuous video dataset is jointly trained by balancing the Connectionist Temporal Classification (CTC) loss and the maximum-entropy loss. The proposed approach can capture better fine-grained attention in a single iteration. Experiment evaluations show that it outperforms the state-of-the-art approaches.
In this work we tackle the task of video-based visual emotion recognition in the wild. Standard methodologies that rely solely on the extraction of bodily and facial features often fall short of accurate emotion prediction in cases where the aforementioned sources of affective information are inaccessible due to head/body orientation, low resolution and poor illumination. We aspire to alleviate this problem by leveraging visual context in the form of scene characteristics and attributes, as part of a broader emotion recognition framework. Temporal Segment Networks (TSN) constitute the backbone of our proposed model. Apart from the RGB input modality, we make use of dense Optical Flow, following an intuitive multi-stream approach for a more effective encoding of motion. Furthermore, we shift our attention towards skeleton-based learning and leverage action-centric data as means of pre-training a Spatial-Temporal Graph Convolutional Network (ST-GCN) for the task of emotion recognition. Our extensive experiments on the challenging Body Language Dataset (BoLD) verify the superiority of our methods over existing approaches, while by properly incorporating all of the aforementioned modules in a network ensemble, we manage to surpass the previous best published recognition scores, by a large margin.
We present SMURF, a method for unsupervised learning of optical flow that improves state of the art on all benchmarks by $36\%$ to $40\%$ (over the prior best method UFlow) and even outperforms several supervised approaches such as PWC-Net and FlowNet2. Our method integrates architecture improvements from supervised optical flow, i.e. the RAFT model, with new ideas for unsupervised learning that include a sequence-aware self-supervision loss, a technique for handling out-of-frame motion, and an approach for learning effectively from multi-frame video data while still only requiring two frames for inference.
Video salient object detection (VSOD) aims to locate and segment the most attractive object by exploiting both spatial cues and temporal cues hidden in video sequences. However, spatial and temporal cues are often unreliable in real-world scenarios, such as low-contrast foreground, fast motion, and multiple moving objects. To address these problems, we propose a new framework to adaptively capture available information from spatial and temporal cues, which contains Confidence-guided Adaptive Gate (CAG) modules and Dual Differential Enhancement (DDE) modules. For both RGB features and optical flow features, CAG estimates confidence scores supervised by the IoU between predictions and the ground truths to re-calibrate the information with a gate mechanism. DDE captures the differential feature representation to enrich the spatial and temporal information and generate the fused features. Experimental results on four widely used datasets demonstrate the effectiveness of the proposed method against thirteen state-of-the-art methods.
Facial expression spotting is the preliminary step for micro- and macro-expression analysis. The task of reliably spotting such expressions in video sequences is currently unsolved. The current best systems depend upon optical flow methods to extract regional motion features, before categorisation of that motion into a specific class of facial movement. Optical flow is susceptible to drift error, which introduces a serious problem for motions with long-term dependencies, such as high frame-rate macro-expression. We propose a purely deep learning solution which, rather than tracking frame differential motion, compares via a convolutional model, each frame with two temporally local reference frames. Reference frames are sampled according to calculated micro- and macro-expression duration. As baseline for MEGC2021 using leave-one-subject-out evaluation method, we show that our solution achieves F1-score of 0.105 in a high frame-rate (200 fps) SAMM long videos dataset (SAMM-LV) and is competitive in a low frame-rate (30 fps) (CAS(ME)2) dataset. On unseen MEGC2022 challenge dataset, the baseline results are 0.1176 on SAMM Challenge dataset, 0.1739 on CAS(ME)3 and overall performance of 0.1531 on both dataset.
Video Frame Interpolation synthesizes non-existent images between adjacent frames, with the aim of providing a smooth and consistent visual experience. Two approaches for solving this challenging task are optical flow based and kernel-based methods. In existing works, optical flow based methods can provide accurate point-to-point motion description, however, they lack constraints on object structure. On the contrary, kernel-based methods focus on structural alignment, which relies on semantic and apparent features, but tends to blur results. Based on these observations, we propose a structure-motion based iterative fusion method. The framework is an end-to-end learnable structure with two stages. First, interpolated frames are synthesized by structure-based and motion-based learning branches respectively, then, an iterative refinement module is established via spatial and temporal feature integration. Inspired by the observation that audiences have different visual preferences on foreground and background objects, we for the first time propose to use saliency masks in the evaluation processes of the task of video frame interpolation. Experimental results on three typical benchmarks show that the proposed method achieves superior performance on all evaluation metrics over the state-of-the-art methods, even when our models are trained with only one-tenth of the data other methods use.
Most Video Super-Resolution (VSR) methods enhance a video reference frame by aligning its neighboring frames and mining information on these frames. Recently, deformable alignment has drawn extensive attention in VSR community for its remarkable performance, which can adaptively align neighboring frames with the reference one. However, we experimentally find that deformable alignment methods still suffer from fast motion due to locally loss-driven offset prediction and lack explicit motion constraints. Hence, we propose a Matching-based Flow Estimation (MFE) module to conduct global semantic feature matching and estimate optical flow as coarse offset for each location. And a Flow-guided Deformable Module (FDM) is proposed to integrate optical flow into deformable convolution. The FDM uses the optical flow to warp the neighboring frames at first. And then, the warped neighboring frames and the reference one are used to predict a set of fine offsets for each coarse offset. In general, we propose an end-to-end deep network called Flow-guided Deformable Alignment Network (FDAN), which reaches the state-of-the-art performance on two benchmark datasets while is still competitive in computation and memory consumption.
With the advent of neuromorphic vision sensors such as event-based cameras, a paradigm shift is required for most computer vision algorithms. Among these algorithms, optical flow estimation is a prime candidate for this process considering that it is linked to a neuromorphic vision approach. Usage of optical flow is widespread in robotics applications due to its richness and accuracy. We present a Principal Component Analysis (PCA) approach to the problem of event-based optical flow estimation. In this approach, we examine different regularization methods which efficiently enhance the estimation of the optical flow. We show that the best variant of our proposed method, dedicated to the real-time context of visual odometry, is about two times faster compared to state-of-the-art implementations while significantly improves optical flow accuracy.
Video anomaly detection is a challenging task because of diverse abnormal events. To this task, methods based on reconstruction and prediction are wildly used in recent works, which are built on the assumption that learning on normal data, anomalies cannot be reconstructed or predicated as good as normal patterns, namely the anomaly result with more errors. In this paper, we propose to discriminate anomalies from normal ones by the duality of normality-granted optical flow, which is conducive to predict normal frames but adverse to abnormal frames. The normality-granted optical flow is predicted from a single frame, to keep the motion knowledge focused on normal patterns. Meanwhile, We extend the appearance-motion correspondence scheme from frame reconstruction to prediction, which not only helps to learn the knowledge about object appearances and correlated motion, but also meets the fact that motion is the transformation between appearances. We also introduce a margin loss to enhance the learning of frame prediction. Experiments on standard benchmark datasets demonstrate the impressive performance of our approach.
Intersections where vehicles are permitted to turn and interact with vulnerable road users (VRUs) like pedestrians and cyclists are among some of the most challenging locations for automated and accurate recognition of road users' behavior. In this paper, we propose a deep conditional generative model for interaction detection at such locations. It aims to automatically analyze massive video data about the continuity of road users' behavior. This task is essential for many intelligent transportation systems such as traffic safety control and self-driving cars that depend on the understanding of road users' locomotion. A Conditional Variational Auto-Encoder based model with Gaussian latent variables is trained to encode road users' behavior and perform probabilistic and diverse predictions of interactions. The model takes as input the information of road users' type, position and motion automatically extracted by a deep learning object detector and optical flow from videos, and generates frame-wise probabilities that represent the dynamics of interactions between a turning vehicle and any VRUs involved. The model's efficacy was validated by testing on real--world datasets acquired from two different intersections. It achieved an F1-score above 0.96 at a right--turn intersection in Germany and 0.89 at a left--turn intersection in Japan, both with very busy traffic flows.
With the goal of predicting the future rainfall intensity in a local region over a relatively short period time, precipitation nowcasting has been a long-time scientific challenge with great social and economic impact. The radar echo extrapolation approaches for precipitation nowcasting take radar echo images as input, aiming to generate future radar echo images by learning from the historical images. To effectively handle complex and high non-stationary evolution of radar echoes, we propose to decompose the movement into optical flow field motion and morphologic deformation. Following this idea, we introduce Flow-Deformation Network (FDNet), a neural network that models flow and deformation in two parallel cross pathways. The flow encoder captures the optical flow field motion between consecutive images and the deformation encoder distinguishes the change of shape from the translational motion of radar echoes. We evaluate the proposed network architecture on two real-world radar echo datasets. Our model achieves state-of-the-art prediction results compared with recent approaches. To the best of our knowledge, this is the first network architecture with flow and deformation separation to model the evolution of radar echoes for precipitation nowcasting. We believe that the general idea of this work could not only inspire much more effective approaches but also be applied to other similar spatiotemporal prediction tasks
The goal of this paper is to self-train a 3D convolutional neural network on an unlabeled video collection for deployment on small-scale video collections. As smaller video datasets benefit more from motion than appearance, we strive to train our network using optical flow, but avoid its computation during inference. We propose the first motion-augmented self-training regime, we call MotionFit. We start with supervised training of a motion model on a small, and labeled, video collection. With the motion model we generate pseudo-labels for a large unlabeled video collection, which enables us to transfer knowledge by learning to predict these pseudo-labels with an appearance model. Moreover, we introduce a multi-clip loss as a simple yet efficient way to improve the quality of the pseudo-labeling, even without additional auxiliary tasks. We also take into consideration the temporal granularity of videos during self-training of the appearance model, which was missed in previous works. As a result we obtain a strong motion-augmented representation model suited for video downstream tasks like action recognition and clip retrieval. On small-scale video datasets, MotionFit outperforms alternatives for knowledge transfer by 5%-8%, video-only self-supervision by 1%-7% and semi-supervised learning by 9%-18% using the same amount of class labels.