In the past few years, we have witnessed remarkable breakthroughs in self-supervised representation learning. Despite the success and adoption of representations learned through this paradigm, much is yet to be understood about how different training methods and datasets influence performance on downstream tasks. In this paper, we analyze contrastive approaches as one of the most successful and popular variants of self-supervised representation learning. We perform this analysis from the perspective of the training algorithms, pre-training datasets and end tasks. We examine over 700 training experiments including 30 encoders, 4 pre-training datasets and 20 diverse downstream tasks. Our experiments address various questions regarding the performance of self-supervised models compared to their supervised counterparts, current benchmarks used for evaluation, and the effect of the pre-training data on end task performance. We hope the insights and empirical evidence provided by this work will help future research in learning better visual representations.
论文检索
输入标题、作者或关键词，从 6,469 篇学术成果中精准定位
Video frame interpolation aims to synthesize accurate intermediate frames given a low-frame-rate video. While the quality of the generated frames is increasingly getting better, state-of-the-art models have become more and more computationally expensive. However, local regions with small or no motion can be easily interpolated with simple models and do not require such heavy compute, whereas some regions may not be correct even after inference through a large model. Thus, we propose an effective framework that assigns varying amounts of computation for different regions. Our dynamic architecture first calculates the approximate motion magnitude to use as a proxy for the difficulty levels for each region, and decides the depth of the model and the scale of the input. Experimental results show that static regions pass through a smaller number of layers, while the regions with larger motion are downscaled for better motion reasoning. In doing so, we demonstrate that the proposed framework can significantly reduce the computation cost (FLOPs) while maintaining the performance, often up to 50% when interpolating a 2K resolution video.
Time-Multiplexed Coded Aperture Imaging: Learned Coded Aperture and Pixel Exposures for Compressive Imaging Systems
PDF ↗Compressive imaging using coded apertures (CA) is a powerful technique that can be used to recover depth, light fields, hyperspectral images and other quantities from a single snapshot. The performance of compressive imaging systems based on CAs mostly depends on two factors: the properties of the mask's attenuation pattern, that we refer to as "codification", and the computational techniques used to recover the quantity of interest from the coded snapshot. In this work, we introduce the idea of using time-varying CAs synchronized with spatially varying pixel shutters. We divide the exposure of a sensor into sub-exposures at the beginning of which the CA mask changes and at which the sensor's pixels are simultaneously and individually switched "on" or "off". This is a practically appealing codification as it does not introduce additional optical components other than the already present CA but uses a change in the pixel shutter that can be easily realized electronically. We show that our proposed time-multiplexed coded aperture (TMCA) can be optimized end to end and induces better coded snapshots enabling superior reconstructions in two different applications: compressive light field imaging and hyperspectral imaging. We demonstrate both in simulation and with real captures (taken with prototypes we built) that this codification outperforms the state-of-the-art compressive imaging systems by a large margin in those applications.
This paper proposes a novel deep learning approach for single image defocus deblurring based on inverse kernels. In a defocused image, the blur shapes are similar among pixels although the blur sizes can spatially vary. To utilize the property with inverse kernels, we exploit the observation that when only the size of a defocus blur changes while keeping the shape, the shape of the corresponding inverse kernel remains the same and only the scale changes. Based on the observation, we propose a kernel-sharing parallel atrous convolutional (KPAC) block specifically designed by incorporating the property of inverse kernels for single image defocus deblurring. To effectively simulate the invariant shapes of inverse kernels with different scales, KPAC shares the same convolutional weights among multiple atrous convolution layers. To efficiently simulate the varying scales of inverse kernels, KPAC consists of only a few atrous convolution layers with different dilations and learns per-pixel scale attentions to aggregate the outputs of the layers. KPAC also utilizes the shape attention to combine the outputs of multiple convolution filters in each atrous convolution layer, to deal with defocus blur with a slightly varying shape. We demonstrate that our approach achieves state-of-the-art performance with a much smaller number of parameters than previous methods.
In this paper we contribute a simple yet effective approach for estimating 3D poses of multiple people from multi-view images. Our proposed coarse-to-fine pipeline first aggregates noisy 2D observations from multiple camera views into 3D space and then associates them into individual instances based on a confidence-aware majority voting technique. The final pose estimates are attained from a novel optimization scheme which links high-confidence multi-view 2D observations and 3D joint candidates. Moreover, a statistical parametric body model such as SMPL is leveraged as a regularizing prior for these 3D joint candidates. Specifically, both 3D poses and SMPL parameters are optimized jointly in an alternating fashion. Here the parametric models help in correcting implausible 3D pose estimates and filling in missing joint detections while updated 3D poses in turn guide obtaining better SMPL estimations. By linking 2D and 3D observations, our method is both accurate and generalizes to different data sources because it better decouples the final 3D pose from the inter-person constellation and is more robust to noisy 2D detections. We systematically evaluate our method on public datasets and achieve state-of-the-art performance. The code and video will be available on the project page: https://ait.ethz.ch/projects/2021/multi-human-pose/.
Contrastive learning allows us to flexibly define powerful losses by contrasting positive pairs from sets of negative samples. Recently, the principle has also been used to learn cross-modal embeddings for video and text, yet without exploiting its full potential. In particular, previous losses do not take the intra-modality similarities into account, which leads to inefficient embeddings, as the same content is mapped to multiple points in the embedding space. With CrossCLR, we present a contrastive loss that fixes this issue. Moreover, we define sets of highly related samples in terms of their input embeddings and exclude them from the negative samples to avoid issues with false negatives. We show that these principles consistently improve the quality of the learned embeddings. The joint embeddings learned with CrossCLR extend the state of the art in video-text retrieval on Youcook2 and LSMDC datasets and in video captioning on the Youcook2 dataset by a large margin. We also demonstrate the generality of the concept by learning improved joint embeddings for other pairs of modalities.
Contrastive learning methods have significantly narrowed the gap between supervised and unsupervised learning on computer vision tasks. In this paper, we explore their application to geo-located datasets, e.g. remote sensing, where unlabeled data is often abundant but labeled data is scarce. We first show that due to their different characteristics, a non-trivial gap persists between contrastive and supervised learning on standard benchmarks. To close the gap, we propose novel training methods that exploit the spatio-temporal structure of remote sensing data. We leverage spatially aligned images over time to construct temporal positive pairs in contrastive learning and geo-location to design pre-text tasks. Our experiments show that our proposed method closes the gap between contrastive and supervised learning on image classification, object detection and semantic segmentation for remote sensing. Moreover, we demonstrate that the proposed method can also be applied to geo-tagged ImageNet images, improving downstream performance on various tasks.
Nowadays, there is an abundance of data involving images and surrounding free-form text weakly corresponding to those images. Weakly Supervised phrase-Grounding (WSG) deals with the task of using this data to learn to localize (or to ground) arbitrary text phrases in images without any additional annotations. However, most recent SotA methods for WSG assume an existence of a pre-trained object detector, relying on it to produce the ROIs for localization. In this work, we focus on the task of Detector-Free WSG (DF-WSG) to solve WSG without relying on a pre-trained detector. We directly learn everything from the images and associated free-form text pairs, thus potentially gaining advantage on the categories unsupported by the detector. The key idea behind our proposed Grounding by Separation (GbS) method is synthesizing `text to image-regions' associations by random alpha-blending of arbitrary image pairs and using the corresponding texts of the pair as conditions to recover the alpha map from the blended image via a segmentation network. At test time, this allows using the query phrase as a condition for a non-blended query image, thus interpreting the test image as a composition of a region corresponding to the phrase and the complement region. Using this approach we demonstrate a significant accuracy improvement, up to 8.5% over previous DF-WSG SotA, for a range of benchmarks including Flickr30K, Visual Genome, and ReferIt, as well as a significant complementary improvement (above 7%) over the detector-based approaches for WSG.
Scene graph generation has emerged as an important problem in computer vision. While scene graphs provide a grounded representation of objects, their locations and relations in an image, they do so only at the granularity of proposal bounding boxes. In this work, we propose the first, to our knowledge, framework for pixel-level segmentation-grounded scene graph generation. Our framework is agnostic to the underlying scene graph generation method and address the lack of segmentation annotations in target scene graph datasets (e.g., Visual Genome) through transfer and multi-task learning from, and with, an auxiliary dataset (e.g., MS COCO). Specifically, each target object being detected is endowed with a segmentation mask, which is expressed as a lingual-similarity weighted linear combination over categories that have annotations present in an auxiliary dataset. These inferred masks, along with a Gaussian masking mechanism which grounds the relations at a pixel-level within the image, allow for improved relation prediction. The entire framework is end-to-end trainable and is learned in a multi-task manner.
Due to the stochasticity of human behaviors, predicting the future trajectories of road agents is challenging for autonomous driving. Recently, goal-based multi-trajectory prediction methods are proved to be effective, where they first score over-sampled goal candidates and then select a final set from them. However, these methods usually involve goal predictions based on sparse pre-defined anchors and heuristic goal selection algorithms. In this work, we propose an anchor-free and end-to-end trajectory prediction model, named DenseTNT, that directly outputs a set of trajectories from dense goal candidates. In addition, we introduce an offline optimization-based technique to provide multi-future pseudo-labels for our final online model. Experiments show that DenseTNT achieves state-of-the-art performance, ranking 1st on the Argoverse motion forecasting benchmark and being the 1st place winner of the 2021 Waymo Open Dataset Motion Prediction Challenge.
Neural trees aim at integrating deep neural networks and decision trees so as to bring the best of the two worlds, including representation learning from the former and faster inference from the latter. In this paper, we introduce a novel approach, termed as Self-born Wiring (SeBoW), to learn neural trees from a mother deep neural network. In contrast to prior neural-tree approaches that either adopt a pre-defined structure or grow hierarchical layers in a progressive manner, task-adaptive neural trees in SeBoW evolve from a deep neural network through a construction-by-destruction process, enabling a global-level parameter optimization that further yields favorable results. Specifically, given a designated network configuration like VGG, SeBoW disconnects all the layers and derives isolated filter groups, based on which a global-level wiring process is conducted to attach a subset of filter groups, eventually bearing a lightweight neural tree. Extensive experiments demonstrate that, with a lower computational cost, SeBoW outperforms all prior neural trees by a significant margin and even achieves results on par with predominant non-tree networks like ResNets. Moreover, SeBoW proves its scalability to large-scale datasets like ImageNet, which has been barely explored by prior tree networks.
Test-time augmentation---the aggregation of predictions across transformed versions of a test input---is a common practice in image classification. Traditionally, predictions are combined using a simple average. In this paper, we present 1) experimental analyses that shed light on cases in which the simple average is suboptimal and 2) a method to address these shortcomings. A key finding is that even when test-time augmentation produces a net improvement in accuracy, it can change many correct predictions into incorrect predictions. We delve into when and why test-time augmentation changes a prediction from being correct to incorrect and vice versa. Building on these insights, we present a learning-based method for aggregating test-time augmentations. Experiments across a diverse set of models, datasets, and augmentations show that our method delivers consistent improvements over existing approaches.
This paper proposes a novel location-aware deep-learning-based single image reflection removal method. Our network has a reflection detection module to regress a probabilistic reflection confidence map, taking multi-scale Laplacian features as inputs. This probabilistic map tells if a region is reflection-dominated or transmission-dominated, and it is used as a cue for the network to control the feature flow when predicting the reflection and transmission layers. We design our network as a recurrent network to progressively refine reflection removal results at each iteration. The novelty is that we leverage Laplacian kernel parameters to emphasize the boundaries of strong reflections. It is beneficial to strong reflection detection and substantially improves the quality of reflection removal results. Extensive experiments verify the superior performance of the proposed method over state-of-the-art approaches. Our code and the pre-trained model can be found at https://github.com/zdlarr/Location-aware-SIRR.
Rapid progress in 3D semantic segmentation is inseparable from the advances of deep network models, which highly rely on large-scale annotated data for training. To address the high cost and challenges of 3D point-level labeling, we present a method for semi-supervised point cloud semantic segmentation to adopt unlabeled point clouds in training to boost the model performance. Inspired by the recent contrastive loss in self-supervised tasks, we propose the guided point contrastive loss to enhance the feature representation and model generalization ability in semi-supervised setting. Semantic predictions on unlabeled point clouds serve as pseudo-label guidance in our loss to avoid negative pairs in the same category. Also, we design the confidence guidance to ensure high-quality feature learning. Besides, a category-balanced sampling strategy is proposed to collect positive and negative samples to mitigate the class imbalance problem. Extensive experiments on three datasets (ScanNet V2, S3DIS, and SemanticKITTI) show the effectiveness of our semi-supervised method to improve the prediction quality with unlabeled data.
Vulnerability of 3D point cloud (PC) classifiers has become a grave concern due to the popularity of 3D sensors in safety-critical applications. Existing adversarial attacks against 3D PC classifiers are all test-time evasion (TTE) attacks that aim to induce test-time misclassifications using knowledge of the classifier. But since the victim classifier is usually not accessible to the attacker, the threat is largely diminished in practice, as PC TTEs typically have poor transferability. Here, we propose the first backdoor attack (BA) against PC classifiers. Originally proposed for images, BAs poison the victim classifier's training set so that the classifier learns to decide to the attacker's target class whenever the attacker's backdoor pattern is present in a given input sample. Significantly, BAs do not require knowledge of the victim classifier. Different from image BAs, we propose to insert a cluster of points into a PC as a robust backdoor pattern customized for 3D PCs. Such clusters are also consistent with a physical attack (i.e., with a captured object in a scene). We optimize the cluster's location using an independently trained surrogate classifier and choose the cluster's local geometry to evade possible PC preprocessing and PC anomaly detectors (ADs). Experimentally, our BA achieves a uniformly high success rate (>=87%) and shows evasiveness against state-of-the-art PC ADs. Code is available at https://github.com/zhenxianglance/PCBA.
Many super-resolution (SR) models are optimized for high performance only and therefore lack efficiency due to large model complexity. As large models are often not practical in real-world applications, we investigate and propose novel loss functions, to enable SR with high perceptual quality from much more efficient models. The representative power for a given low-complexity generator network can only be fully leveraged by strong guidance towards the optimal set of parameters. We show that it is possible to improve the performance of a recently introduced efficient generator architecture solely with the application of our proposed loss functions. In particular, we use a Fourier space supervision loss for improved restoration of missing high-frequency (HF) content from the ground truth image and design a discriminator architecture working directly in the Fourier domain to better match the target HF distribution. We show that our losses' direct emphasis on the frequencies in Fourier-space significantly boosts the perceptual image quality, while at the same time retaining high restoration quality in comparison to previously proposed loss functions for this task. The performance is further improved by utilizing a combination of spatial and frequency domain losses, as both representations provide complementary information during training. On top of that, the trained generator achieves comparable results with and is 2.4x and 48x faster than state-of-the-art perceptual SR methods RankSRGAN and SRFlow respectively.
In this paper, we focus on recognizing 3D shapes from arbitrary views, i.e., arbitrary numbers and positions of viewpoints. It is a challenging and realistic setting for view-based 3D shape recognition. We propose a canonical view representation to tackle this challenge. We first transform the original features of arbitrary views to a fixed number of view features, dubbed canonical view representation, by aligning the arbitrary view features to a set of learnable reference view features using optimal transport. In this way, each 3D shape with arbitrary views is represented by a fixed number of canonical view features, which are further aggregated to generate a rich and robust 3D shape representation for shape recognition. We also propose a canonical view feature separation constraint to enforce that the view features in canonical view representation can be embedded into scattered points in a Euclidean space. Experiments on the ModelNet40, ScanObjectNN, and RGBD datasets show that our method achieves competitive results under the fixed viewpoint settings, and significantly outperforms the applicable methods under the arbitrary view setting.
Deep Convolutional Neural Networks (CNNs) have long been the architecture of choice for computer vision tasks. Recently, Transformer-based architectures like Vision Transformer (ViT) have matched or even surpassed ResNets for image classification. However, details of the Transformer architecture such as the use of non-overlapping patches lead one to wonder whether these networks are as robust. In this paper, we perform an extensive study of a variety of different measures of robustness of ViT models and compare the findings to ResNet baselines. We investigate robustness to input perturbations as well as robustness to model perturbations. We find that when pre-trained with a sufficient amount of data, ViT models are at least as robust as the ResNet counterparts on a broad range of perturbations. We also find that Transformers are robust to the removal of almost any single layer, and that while activations from later layers are highly correlated with each other, they nevertheless play an important role in classification.
Deep generative models of 3D shapes have received a great deal of research interest. Yet, almost all of them generate discrete shape representations, such as voxels, point clouds, and polygon meshes. We present the first 3D generative model for a drastically different shape representation --- describing a shape as a sequence of computer-aided design (CAD) operations. Unlike meshes and point clouds, CAD models encode the user creation process of 3D shapes, widely used in numerous industrial and engineering design tasks. However, the sequential and irregular structure of CAD operations poses significant challenges for existing 3D generative models. Drawing an analogy between CAD operations and natural language, we propose a CAD generative network based on the Transformer. We demonstrate the performance of our model for both shape autoencoding and random shape generation. To train our network, we create a new CAD dataset consisting of 178,238 models and their CAD construction sequences. We have made this dataset publicly available to promote future research on this topic.
We present a method for reconstructing the 3D shape of underwater environments from a single, stationary camera placed above the water. We propose a novel differentiable framework, which, to our knowledge, is the first single-camera solution that is capable of simultaneously retrieving the structure of dynamic water surfaces and static underwater scene geometry in the wild. This framework integrates ray casting of Snell's law at the refractive interface, multi-view triangulation and specially designed loss functions. Our method is calibration-free, and thus it is easy to collect data outdoors in uncontrolled environments. Experimental results show that our method is able to realize robust and quality reconstructions on a variety of scenes, both in a laboratory environment and in the wild, and even in a salt water environment. We believe the method is promising for applications in surveying and environmental monitoring.