Fisheye cameras are commonly used in applications like autonomous driving and surveillance to provide a large field of view ($>180^{\circ}$). However, they come at the cost of strong non-linear distortions which require more complex algorithms. In this paper, we explore Euclidean distance estimation on fisheye cameras for automotive scenes. Obtaining accurate and dense depth supervision is difficult in practice, but self-supervised learning approaches show promising results and could potentially overcome the problem. We present a novel self-supervised scale-aware framework for learning Euclidean distance and ego-motion from raw monocular fisheye videos without applying rectification. While it is possible to perform piece-wise linear approximation of fisheye projection surface and apply standard rectilinear models, it has its own set of issues like re-sampling distortion and discontinuities in transition regions. To encourage further research in this area, we will release our dataset as part of the WoodScape project \cite{yogamani2019woodscape}. We further evaluated the proposed algorithm on the KITTI dataset and obtained state-of-the-art results comparable to other self-supervised monocular methods. Qualitative results on an unseen fisheye video demonstrate impressive performance https://youtu.be/Sgq1WzoOmXg.
Although deep networks have significantly increased the performance of visual recognition methods, it is still challenging to achieve the robustness across visual domains that is necessary for real-world applications. To tackle this issue, research on domain adaptation and generalization has flourished over the last decade. An important aspect to consider when assessing the work done in the literature so far is the amount of data annotation necessary for training each approach, both at the source and target level. In this paper we argue that the data annotation overload should be minimal, as it is costly. Hence, we propose to use self-supervised learning to achieve domain generalization and adaptation. We consider learning regularities from non annotated data as an auxiliary task, and cast the problem within an Auxiliary Learning principled framework. Moreover, we suggest to further exploit the ability to learn about visual domains from non annotated images by learning from target data while testing, as data are presented to the algorithm one sample at a time. Results on three different scenarios confirm the value of our approach.
This paper introduces a challenging object grasping task and proposes a self-supervised learning approach. The goal of the task is to grasp an object which is not feasible with a single parallel gripper, but only with harnessing environment fixtures (e.g., walls, furniture, heavy objects). This Slide-to-Wall grasping task assumes no prior knowledge except the partial observation of a target object. Hence the robot should learn an effective policy given a scene observation that may include the target object, environmental fixtures, and any other disturbing objects. We formulate the problem as visual affordances learning for which Target-Oriented Deep Q-Network (TO-DQN) is proposed to efficiently learn visual affordance maps (i.e., Q-maps) to guide robot actions. Since the training necessitates robot's exploration and collision with the fixtures, TO-DQN is first trained safely with a simulated robot manipulator and then applied to a real robot. We empirically show that TO-DQN can learn to solve the task in different environment settings in simulation and outperforms a standard and a variant of Deep Q-Network (DQN) in terms of training efficiency and robustness. The testing performance in both simulation and real-robot experiments shows that the policy trained by TO-DQN achieves comparable performance to humans.
We investigate the role of self-supervised learning (SSL) in the context of few-shot learning. Although recent research has shown the benefits of SSL on large unlabeled datasets, its utility on small datasets is relatively unexplored. We find that SSL reduces the relative error rate of few-shot meta-learners by 4%-27%, even when the datasets are small and only utilizing images within the datasets. The improvements are greater when the training set is smaller or the task is more challenging. Although the benefits of SSL may increase with larger training sets, we observe that SSL can hurt the performance when the distributions of images used for meta-learning and SSL are different. We conduct a systematic study by varying the degree of domain shift and analyzing the performance of several meta-learners on a multitude of domains. Based on this analysis we present a technique that automatically selects images for SSL from a large, generic pool of unlabeled images for a given dataset that provides further improvements.
Recently, self-supervised learning has proved to be effective to learn representations of events suitable for temporal segmentation in image sequences, where events are understood as sets of temporally adjacent images that are semantically perceived as a whole. However, although this approach does not require expensive manual annotations, it is data hungry and suffers from domain adaptation problems. As an alternative, in this work, we propose a novel approach for learning event representations named Dynamic Graph Embedding (DGE). The assumption underlying our model is that a sequence of images can be represented by a graph that encodes both semantic and temporal similarity. The key novelty of DGE is to learn jointly the graph and its graph embedding. At its core, DGE works by iterating over two steps: 1) updating the graph representing the semantic and temporal similarity of the data based on the current data representation, and 2) updating the data representation to take into account the current data graph structure. The main advantage of DGE over state-of-the-art self-supervised approaches is that it does not require any training set, but instead learns iteratively from the data itself a low-dimensional embedding that reflects their temporal and semantic similarity. Experimental results on two benchmark datasets of real image sequences captured at regular time intervals demonstrate that the proposed DGE leads to event representations effective for temporal segmentation. In particular, it achieves robust temporal segmentation on the EDUBSeg and EDUBSeg-Desc benchmark datasets, outperforming the state of the art. Additional experiments on two Human Motion Segmentation benchmark datasets demonstrate the generalization capabilities of the proposed DGE.
Witnessed the development of deep learning, increasing number of studies try to build computer aided diagnosis systems for 3D volumetric medical data. However, as the annotations of 3D medical data are difficult to acquire, the number of annotated 3D medical images is often not enough to well train the deep learning networks. The self-supervised learning deeply exploiting the information of raw data is one of the potential solutions to loose the requirement of training data. In this paper, we propose a self-supervised learning framework for the volumetric medical images. A novel proxy task, i.e., Rubik's cube recovery, is formulated to pre-train 3D neural networks. The proxy task involves two operations, i.e., cube rearrangement and cube rotation, which enforce networks to learn translational and rotational invariant features from raw 3D data. Compared to the train-from-scratch strategy, fine-tuning from the pre-trained network leads to a better accuracy on various tasks, e.g., brain hemorrhage classification and brain tumor segmentation. We show that our self-supervised learning approach can substantially boost the accuracies of 3D deep learning networks on the volumetric medical datasets without using extra data. To our best knowledge, this is the first work focusing on the self-supervised learning of 3D neural networks.
Nowadays, supervised deep learning techniques yield the best state-of-the-art prediction performances for a wide variety of computer vision tasks. However, such supervised techniques generally require a large amount of manually labeled training data. In the context of autonomous vehicles perception, this requirement is critical, as the distribution of sensor data can continuously change and include several unexpected variations. It turns out that a category of learning techniques, referred to as self-supervised learning (SSL), consists of replacing the manual labeling effort by an automatic labeling process. Thanks to their ability to learn on the application time and in varying environments, state-of-the-art SSL techniques provide a valid alternative to supervised learning for a variety of different tasks, including long-range traversable area segmentation, moving obstacle instance segmentation, long-term moving obstacle tracking, or depth map prediction. In this tutorial-style article, we present an overview and a general formalization of the concept of self-supervised learning (SSL) for autonomous vehicles perception. This formalization provides helpful guidelines for developing novel frameworks based on generic SSL principles. Moreover, it enables to point out significant challenges in the design of future SSL systems.
In this paper, we propose Test-Time Training, a general approach for improving the performance of predictive models when training and test data come from different distributions. We turn a single unlabeled test sample into a self-supervised learning problem, on which we update the model parameters before making a prediction. This also extends naturally to data in an online stream. Our simple approach leads to improvements on diverse image classification benchmarks aimed at evaluating robustness to distribution shifts.
Learning to predict scene depth and camera motion from RGB inputs only is a challenging task. Most existing learning based methods deal with this task in a supervised manner which require ground-truth data that is expensive to acquire. More recent approaches explore the possibility of estimating scene depth and camera pose in a self-supervised learning framework. Despite encouraging results are shown, current methods either learn from monocular videos for depth and pose and typically do so without enforcing multi-view geometry constraints between scene structure and camera motion, or require stereo sequences as input where the ground-truth between-frame motion parameters need to be known. In this paper we propose to jointly optimize the scene depth and camera motion via incorporating differentiable Bundle Adjustment (BA) layer by minimizing the feature-metric error, and then form the photometric consistency loss with view synthesis as the final supervisory signal. The proposed approach only needs unlabeled monocular videos as input, and extensive experiments on the KITTI and Cityscapes dataset show that our method achieves state-of-the-art results in self-supervised approaches using monocular videos as input, and even gains advantage to the line of methods that learns from calibrated stereo sequences (i.e. with pose supervision).
Most of the recent Deep Semantic Segmentation algorithms suffer from large generalization errors, even when powerful hierarchical representation models based on convolutional neural networks have been employed. This could be attributed to limited training data and large distribution gap in train and test domain datasets. In this paper, we propose a multi-level self-supervised learning model for domain adaptation of semantic segmentation. Exploiting the idea that an object (and most of the stuff given context) should be labeled consistently regardless of its location, we generate spatially independent and semantically consistent (SISC) pseudo-labels by segmenting multiple sub-images using base model and designing an aggregation strategy. Image level pseudo weak-labels, PWL, are computed to guide domain adaptation by capturing global context similarity in source and domain at latent space level. Thus helping latent space learn the representation even when there are very few pixels belonging to the domain category (small object for example) compared to rest of the image. Our multi-level Self-supervised learning (MLSL) outperforms existing state-of art (self or adversarial learning) algorithms. Specifically, keeping all setting similar and employing MLSL we obtain an mIoU gain of 5:1% on GTA-V to Cityscapes adaptation and 4:3% on SYNTHIA to Cityscapes adaptation compared to existing state-of-art method.
This paper proposes to learn reliable dense correspondence from videos in a self-supervised manner. Our learning process integrates two highly related tasks: tracking large image regions \emph{and} establishing fine-grained pixel-level associations between consecutive video frames. We exploit the synergy between both tasks through a shared inter-frame affinity matrix, which simultaneously models transitions between video frames at both the region- and pixel-levels. While region-level localization helps reduce ambiguities in fine-grained matching by narrowing down search regions; fine-grained matching provides bottom-up features to facilitate region-level localization. Our method outperforms the state-of-the-art self-supervised methods on a variety of visual correspondence tasks, including video-object and part-segmentation propagation, keypoint tracking, and object tracking. Our self-supervised method even surpasses the fully-supervised affinity feature representation obtained from a ResNet-18 pre-trained on the ImageNet.
Increasing model size when pretraining natural language representations often results in improved performance on downstream tasks. However, at some point further model increases become harder due to GPU/TPU memory limitations and longer training times. To address these problems, we present two parameter-reduction techniques to lower memory consumption and increase the training speed of BERT. Comprehensive empirical evidence shows that our proposed methods lead to models that scale much better compared to the original BERT. We also use a self-supervised loss that focuses on modeling inter-sentence coherence, and show it consistently helps downstream tasks with multi-sentence inputs. As a result, our best model establishes new state-of-the-art results on the GLUE, RACE, and \squad benchmarks while having fewer parameters compared to BERT-large. The code and the pretrained models are available at https://github.com/google-research/ALBERT.
The self-supervised learning of depth and pose from monocular sequences provides an attractive solution by using the photometric consistency of nearby frames as it depends much less on the ground-truth data. In this paper, we address the issue when previous assumptions of the self-supervised approaches are violated due to the dynamic nature of real-world scenes. Different from handling the noise as uncertainty, our key idea is to incorporate more robust geometric quantities and enforce internal consistency in the temporal image sequence. As demonstrated on commonly used benchmark datasets, the proposed method substantially improves the state-of-the-art methods on both depth and relative pose estimation for monocular image sequences, without adding inference overhead.
Monocular depth estimators can be trained with various forms of self-supervision from binocular-stereo data to circumvent the need for high-quality laser scans or other ground-truth data. The disadvantage, however, is that the photometric reprojection losses used with self-supervised learning typically have multiple local minima. These plausible-looking alternatives to ground truth can restrict what a regression network learns, causing it to predict depth maps of limited quality. As one prominent example, depth discontinuities around thin structures are often incorrectly estimated by current state-of-the-art methods. Here, we study the problem of ambiguous reprojections in depth prediction from stereo-based self-supervision, and introduce Depth Hints to alleviate their effects. Depth Hints are complementary depth suggestions obtained from simple off-the-shelf stereo algorithms. These hints enhance an existing photometric loss function, and are used to guide a network to learn better weights. They require no additional data, and are assumed to be right only sometimes. We show that using our Depth Hints gives a substantial boost when training several leading self-supervised-from-stereo models, not just our own. Further, combined with other good practices, we produce state-of-the-art depth predictions on the KITTI benchmark.
In this paper, we study the problem of how to defend classifiers against adversarial attacks that fool the classifiers using subtly modified input data. In contrast to previous works, here we focus on the white-box adversarial defense where the attackers are granted full access to not only the classifiers but also defenders to produce as strong attacks as possible. In such a context we propose viewing a defender as a functional, a higher-order function that takes functions as its argument to represent a function space, rather than fixed functions conventionally. From this perspective, a defender should be realized and optimized individually for each adversarial input. To this end, we propose RIDE, an efficient and provably convergent self-supervised learning algorithm for individual data estimation to protect the predictions from adversarial attacks. We demonstrate the significant improvement of adversarial defense performance on image recognition, eg, 98%, 76%, 43% test accuracy on MNIST, CIFAR-10, and ImageNet datasets respectively under the state-of-the-art BPDA attacker.
Video prediction models combined with planning algorithms have shown promise in enabling robots to learn to perform many vision-based tasks through only self-supervision, reaching novel goals in cluttered scenes with unseen objects. However, due to the compounding uncertainty in long horizon video prediction and poor scalability of sampling-based planning optimizers, one significant limitation of these approaches is the ability to plan over long horizons to reach distant goals. To that end, we propose a framework for subgoal generation and planning, hierarchical visual foresight (HVF), which generates subgoal images conditioned on a goal image, and uses them for planning. The subgoal images are directly optimized to decompose the task into easy to plan segments, and as a result, we observe that the method naturally identifies semantically meaningful states as subgoals. Across three out of four simulated vision-based manipulation tasks, we find that our method achieves nearly a 200% performance improvement over planning without subgoals and model-free RL approaches. Further, our experiments illustrate that our approach extends to real, cluttered visual scenes. Project page: https://sites.google.com/stanford.edu/hvf
The objective of this paper is self-supervised learning of spatio-temporal embeddings from video, suitable for human action recognition. We make three contributions: First, we introduce the Dense Predictive Coding (DPC) framework for self-supervised representation learning on videos. This learns a dense encoding of spatio-temporal blocks by recurrently predicting future representations; Second, we propose a curriculum training scheme to predict further into the future with progressively less temporal context. This encourages the model to only encode slowly varying spatial-temporal signals, therefore leading to semantic representations; Third, we evaluate the approach by first training the DPC model on the Kinetics-400 dataset with self-supervised learning, and then finetuning the representation on a downstream task, i.e. action recognition. With single stream (RGB only), DPC pretrained representations achieve state-of-the-art self-supervised performance on both UCF101(75.7% top1 acc) and HMDB51(35.7% top1 acc), outperforming all previous learning methods by a significant margin, and approaching the performance of a baseline pre-trained on ImageNet.
We present a self-supervised learning-based pipeline for dense 3D reconstruction from full-length monocular endoscopic videos without a priori modeling of anatomy or shading. Our method only relies on unlabeled monocular endoscopic videos and conventional multi-view stereo algorithms, and requires neither manual interaction nor patient CT in both training and application phases. In a cross-patient study using CT scans as groundtruth, we show that our method is able to produce photo-realistic dense 3D reconstructions with submillimeter mean residual errors from endoscopic videos from unseen patients and scopes.
Blur detection aims at segmenting the blurred areas of a given image. Recent deep learning-based methods approach this problem by learning an end-to-end mapping between the blurred input and a binary mask representing the localization of its blurred areas. Nevertheless, the effectiveness of such deep models is limited due to the scarcity of datasets annotated in terms of blur segmentation, as blur annotation is labour intensive. In this work, we bypass the need for such annotated datasets for end-to-end learning, and instead rely on object proposals and a model for blur generation in order to produce a dataset of synthetically blurred images. This allows us to perform self-supervised learning over the generated image and ground truth blur mask pairs using CNNs, defining a framework that can be employed in purely self-supervised, weakly supervised or semi-supervised configurations. Interestingly, experimental results of such setups over the largest blur segmentation datasets available show that this approach achieves state of the art results in blur segmentation, even without ever observing any real blurred image.
We propose a self-supervised learning framework for visual odometry (VO) that incorporates correlation of consecutive frames and takes advantage of adversarial learning. Previous methods tackle self-supervised VO as a local structure from motion (SfM) problem that recovers depth from single image and relative poses from image pairs by minimizing photometric loss between warped and captured images. As single-view depth estimation is an ill-posed problem, and photometric loss is incapable of discriminating distortion artifacts of warped images, the estimated depth is vague and pose is inaccurate. In contrast to previous methods, our framework learns a compact representation of frame-to-frame correlation, which is updated by incorporating sequential information. The updated representation is used for depth estimation. Besides, we tackle VO as a self-supervised image generation task and take advantage of Generative Adversarial Networks (GAN). The generator learns to estimate depth and pose to generate a warped target image. The discriminator evaluates the quality of generated image with high-level structural perception that overcomes the problem of pixel-wise loss in previous methods. Experiments on KITTI and Cityscapes datasets show that our method obtains more accurate depth with details preserved and predicted pose outperforms state-of-the-art self-supervised methods significantly.