Jan-Michael Frahm

Jialiang Wang, Daniel Scharstein, Akash Bapat et al.

We present the design of a productionized end-to-end stereo depth sensing system that does pre-processing, online stereo rectification, and stereo depth estimation with a fallback to monocular depth estimation when rectification is unreliable. The output of our depth sensing system is then used in a novel view generation pipeline to create 3D computational photography effects using point-of-view images captured by smart glasses. All these steps are executed on-device on the stringent compute budget of a mobile phone, and because we expect the users can use a wide range of smartphones, our design needs to be general and cannot be dependent on a particular hardware or ML accelerator such as a smartphone GPU. Although each of these steps is well studied, a description of a practical system is still lacking. For such a system, all these steps need to work in tandem with one another and fallback gracefully on failures within the system or less than ideal input data. We show how we handle unforeseen changes to calibration, e.g., due to heat, robustly support depth estimation in the wild, and still abide by the memory and latency constraints required for a smooth user experience. We show that our trained models are fast, and run in less than 1s on a six-year-old Samsung Galaxy S8 phone's CPU. Our models generalize well to unseen data and achieve good results on Middlebury and in-the-wild images captured from the smart glasses.

Shuxian Wang, Yubo Zhang, Sarah K. McGill et al.

Reconstructing a 3D surface from colonoscopy video is challenging due to illumination and reflectivity variation in the video frame that can cause defective shape predictions. Aiming to overcome this challenge, we utilize the characteristics of surface normal vectors and develop a two-step neural framework that significantly improves the colonoscopy reconstruction quality. The normal-based depth initialization network trained with self-supervised normal consistency loss provides depth map initialization to the normal-depth refinement module, which utilizes the relationship between illumination and surface normals to refine the frame-wise normal and depth predictions recursively. Our framework's depth accuracy performance on phantom colonoscopy data demonstrates the value of exploiting the surface normals in colonoscopy reconstruction, especially on en face views. Due to its low depth error, the prediction result from our framework will require limited post-processing to be clinically applicable for real-time colonoscopy reconstruction.

Thanh Vu, Baochen Sun, Bodi Yuan et al.

The success of data mixing augmentations in image classification tasks has been well-received. However, these techniques cannot be readily applied to object detection due to challenges such as spatial misalignment, foreground/background distinction, and plurality of instances. To tackle these issues, we first introduce a novel conceptual framework called Supervision Interpolation (SI), which offers a fresh perspective on interpolation-based augmentations by relaxing and generalizing Mixup. Based on SI, we propose LossMix, a simple yet versatile and effective regularization that enhances the performance and robustness of object detectors and more. Our key insight is that we can effectively regularize the training on mixed data by interpolating their loss errors instead of ground truth labels. Empirical results on the PASCAL VOC and MS COCO datasets demonstrate that LossMix can consistently outperform state-of-the-art methods widely adopted for detection. Furthermore, by jointly leveraging LossMix with unsupervised domain adaptation, we successfully improve existing approaches and set a new state of the art for cross-domain object detection.

Thanh Vu, Yanqi Zhou, Chunfeng Wen et al.

In this work, we propose a novel and scalable solution to address the challenges of developing efficient dense predictions on edge platforms. Our first key insight is that MultiTask Learning (MTL) and hardware-aware Neural Architecture Search (NAS) can work in synergy to greatly benefit on-device Dense Predictions (DP). Empirical results reveal that the joint learning of the two paradigms is surprisingly effective at improving DP accuracy, achieving superior performance over both the transfer learning of single-task NAS and prior state-of-the-art approaches in MTL, all with just 1/10th of the computation. To the best of our knowledge, our framework, named EDNAS, is the first to successfully leverage the synergistic relationship of NAS and MTL for DP. Our second key insight is that the standard depth training for multi-task DP can cause significant instability and noise to MTL evaluation. Instead, we propose JAReD, an improved, easy-to-adopt Joint Absolute-Relative Depth loss, that reduces up to 88% of the undesired noise while simultaneously boosting accuracy. We conduct extensive evaluations on standard datasets, benchmark against strong baselines and state-of-the-art approaches, as well as provide an analysis of the discovered optimal architectures.

Jisan Mahmud, Jan-Michael Frahm

We introduce a unified single and multi-view neural implicit 3D reconstruction framework VPFusion. VPFusion attains high-quality reconstruction using both - 3D feature volume to capture 3D-structure-aware context, and pixel-aligned image features to capture fine local detail. Existing approaches use RNN, feature pooling, or attention computed independently in each view for multi-view fusion. RNNs suffer from long-term memory loss and permutation variance, while feature pooling or independently computed attention leads to representation in each view being unaware of other views before the final pooling step. In contrast, we show improved multi-view feature fusion by establishing transformer-based pairwise view association. In particular, we propose a novel interleaved 3D reasoning and pairwise view association architecture for feature volume fusion across different views. Using this structure-aware and multi-view-aware feature volume, we show improved 3D reconstruction performance compared to existing methods. VPFusion improves the reconstruction quality further by also incorporating pixel-aligned local image features to capture fine detail. We verify the effectiveness of VPFusion on the ShapeNet and ModelNet datasets, where we outperform or perform on-par the state-of-the-art single and multi-view 3D shape reconstruction methods.

John Lim, Jan-Michael Frahm, Fabian Monrose

Keystroke inference attacks are a form of side-channel attacks in which an attacker leverages various techniques to recover a user's keystrokes as she inputs information into some display (e.g., while sending a text message or entering her pin). Typically, these attacks leverage machine learning approaches, but assessing the realism of the threat space has lagged behind the pace of machine learning advancements, due in-part, to the challenges in curating large real-life datasets. We aim to overcome the challenge of having limited number of real data by introducing a video domain adaptation technique that is able to leverage synthetic data through supervised disentangled learning. Specifically, for a given domain, we decompose the observed data into two factors of variation: Style and Content. Doing so provides four learned representations: real-life style, synthetic style, real-life content and synthetic content. Then, we combine them into feature representations from all combinations of style-content pairings across domains, and train a model on these combined representations to classify the content (i.e., labels) of a given datapoint in the style of another domain. We evaluate our method on real-life data using a variety of metrics to quantify the amount of information an attacker is able to recover. We show that our method prevents our model from overfitting to a small real-life training set, indicating that our method is an effective form of data augmentation, thereby making keystroke inference attacks more practical.

John Lim, True Price, Fabian Monrose et al.

A vision-based keystroke inference attack is a side-channel attack in which an attacker uses an optical device to record users on their mobile devices and infer their keystrokes. The threat space for these attacks has been studied in the past, but we argue that the defining characteristics for this threat space, namely the strength of the attacker, are outdated. Previous works do not study adversaries with vision systems that have been trained with deep neural networks because these models require large amounts of training data and curating such a dataset is expensive. To address this, we create a large-scale synthetic dataset to simulate the attack scenario for a keystroke inference attack. We show that first pre-training on synthetic data, followed by adopting transfer learning techniques on real-life data, increases the performance of our deep learning models. This indicates that these models are able to learn rich, meaningful representations from our synthetic data and that training on the synthetic data can help overcome the issue of having small, real-life datasets for vision-based key stroke inference attacks. For this work, we focus on single keypress classification where the input is a frame of a keypress and the output is a predicted key. We are able to get an accuracy of 95.6% after pre-training a CNN on our synthetic data and training on a small set of real-life data in an adversarial domain adaptation framework. Source Code for Simulator: https://github.com/jlim13/keystroke-inference-attack-synthetic-dataset-generator-

Noah Frahm, Dongxu Zhao, Andrea Dunn Beltran et al.

Next Best View (NBV) algorithms aim to maximize 3D scene acquisition quality using minimal resources, e.g. number of acquisitions, time taken, or distance traversed. Prior methods often rely on coverage maximization as a proxy for reconstruction quality, but for complex scenes with occlusions and finer details, this is not always sufficient and leads to poor reconstructions. Our key insight is to train an acquisition policy that directly optimizes for reconstruction quality rather than just coverage. To achieve this, we introduce the View Introspection Network (VIN): a lightweight neural network that predicts the Relative Reconstruction Improvement (RRI) of a potential next viewpoint without making any new acquisitions. We use this network to power a simple, yet effective, sequential samplingbased greedy NBV policy. Our approach, VIN-NBV, generalizes to unseen object categories, operates without prior scene knowledge, is adaptable to resource constraints, and can handle occlusions. We show that our RRI fitness criterion leads to a ~30% gain in reconstruction quality over a coverage-based criterion using the same greedy strategy. Furthermore, VIN-NBV also outperforms deep reinforcement learning methods, Scan-RL and GenNBV, by ~40%.

Dongxu Zhao, Daniel Lichy, Pierre-Nicolas Perrin et al.

We propose a fast and generalizable solution to Multi-view Photometric Stereo (MVPS), called MVPSNet. The key to our approach is a feature extraction network that effectively combines images from the same view captured under multiple lighting conditions to extract geometric features from shading cues for stereo matching. We demonstrate these features, termed `Light Aggregated Feature Maps' (LAFM), are effective for feature matching even in textureless regions, where traditional multi-view stereo methods fail. Our method produces similar reconstruction results to PS-NeRF, a state-of-the-art MVPS method that optimizes a neural network per-scene, while being 411$\times$ faster (105 seconds vs. 12 hours) in inference. Additionally, we introduce a new synthetic dataset for MVPS, sMVPS, which is shown to be effective to train a generalizable MVPS method.

Yubo Zhang, Jan-Michael Frahm, Samuel Ehrenstein et al.

One of the key elements of reconstructing a 3D mesh from a monocular video is generating every frame's depth map. However, in the application of colonoscopy video reconstruction, producing good-quality depth estimation is challenging. Neural networks can be easily fooled by photometric distractions or fail to capture the complex shape of the colon surface, predicting defective shapes that result in broken meshes. Aiming to fundamentally improve the depth estimation quality for colonoscopy 3D reconstruction, in this work we have designed a set of training losses to deal with the special challenges of colonoscopy data. For better training, a set of geometric consistency objectives was developed, using both depth and surface normal information. Also, the classic photometric loss was extended with feature matching to compensate for illumination noise. With the training losses powerful enough, our self-supervised framework named ColDE is able to produce better depth maps of colonoscopy data as compared to the previous work utilizing prior depth knowledge. Used in reconstruction, our network is able to reconstruct good-quality colon meshes in real-time without any post-processing, making it the first to be clinically applicable.

Thanh Vu, Marc Eder, True Price et al.

Despite remarkable improvements in speed and accuracy, convolutional neural networks (CNNs) still typically operate as monolithic entities at inference time. This poses a challenge for resource-constrained practical applications, where both computational budgets and performance needs can vary with the situation. To address these constraints, we propose the Any-Width Network (AWN), an adjustable-width CNN architecture and associated training routine that allow for fine-grained control over speed and accuracy during inference. Our key innovation is the use of lower-triangular weight matrices which explicitly address width-varying batch statistics while being naturally suited for multi-width operations. We also show that this design facilitates an efficient training routine based on random width sampling. We empirically demonstrate that our proposed AWNs compare favorably to existing methods while providing maximally granular control during inference.

Johannes Kopf, Kevin Matzen, Suhib Alsisan et al.

3D photography is a new medium that allows viewers to more fully experience a captured moment. In this work, we refer to a 3D photo as one that displays parallax induced by moving the viewpoint (as opposed to a stereo pair with a fixed viewpoint). 3D photos are static in time, like traditional photos, but are displayed with interactive parallax on mobile or desktop screens, as well as on Virtual Reality devices, where viewing it also includes stereo. We present an end-to-end system for creating and viewing 3D photos, and the algorithmic and design choices therein. Our 3D photos are captured in a single shot and processed directly on a mobile device. The method starts by estimating depth from the 2D input image using a new monocular depth estimation network that is optimized for mobile devices. It performs competitively to the state-of-the-art, but has lower latency and peak memory consumption and uses an order of magnitude fewer parameters. The resulting depth is lifted to a layered depth image, and new geometry is synthesized in parallax regions. We synthesize color texture and structures in the parallax regions as well, using an inpainting network, also optimized for mobile devices, on the LDI directly. Finally, we convert the result into a mesh-based representation that can be efficiently transmitted and rendered even on low-end devices and over poor network connections. Altogether, the processing takes just a few seconds on a mobile device, and the result can be instantly viewed and shared. We perform extensive quantitative evaluation to validate our system and compare its new components against the current state-of-the-art.

Aleksander Holynski, David Geraghty, Jan-Michael Frahm et al.

Low-frequency long-range errors (drift) are an endemic problem in 3D structure from motion, and can often hamper reasonable reconstructions of the scene. In this paper, we present a method to dramatically reduce scale and positional drift by using extended structural features such as planes and vanishing points. Unlike traditional feature matches, our extended features are able to span non-overlapping input images, and hence provide long-range constraints on the scale and shape of the reconstruction. We add these features as additional constraints to a state-of-the-art global structure from motion algorithm and demonstrate that the added constraints enable the reconstruction of particularly drift-prone sequences such as long, low field-of-view videos without inertial measurements. Additionally, we provide an analysis of the drift-reducing capabilities of these constraints by evaluating on a synthetic dataset. Our structural features are able to significantly reduce drift for scenes that contain long-spanning man-made structures, such as aligned rows of windows or planar building facades.

Marc Eder, Mykhailo Shvets, John Lim et al.

In this work, we propose "tangent images," a spherical image representation that facilitates transferable and scalable $360^\circ$ computer vision. Inspired by techniques in cartography and computer graphics, we render a spherical image to a set of distortion-mitigated, locally-planar image grids tangent to a subdivided icosahedron. By varying the resolution of these grids independently of the subdivision level, we can effectively represent high resolution spherical images while still benefiting from the low-distortion icosahedral spherical approximation. We show that training standard convolutional neural networks on tangent images compares favorably to the many specialized spherical convolutional kernels that have been developed, while also scaling efficiently to handle significantly higher spherical resolutions. Furthermore, because our approach does not require specialized kernels, we show that we can transfer networks trained on perspective images to spherical data without fine-tuning and with limited performance drop-off. Finally, we demonstrate that tangent images can be used to improve the quality of sparse feature detection on spherical images, illustrating its usefulness for traditional computer vision tasks like structure-from-motion and SLAM.

Jisan Mahmud, Rajat Vikram Singh, Peri Akiva et al.

The rapid development of inexpensive commodity depth sensors has made keypoint detection and matching in the depth image modality an important problem in computer vision. Despite great improvements in recent RGB local feature learning methods, adapting them directly in the depth modality leads to unsatisfactory performance. Most of these methods do not explicitly reason beyond the visible pixels in the images. To address the limitations of these methods, we propose a framework ViewSynth, to jointly learn: (1) viewpoint invariant keypoint-descriptor from depth images using a proposed Contrastive Matching Loss, and (2) view synthesis of depth images from different viewpoints using the proposed View Synthesis Module and View Synthesis Loss. By learning view synthesis, we explicitly encourage the feature extractor to encode information about not only the visible, but also the occluded parts of the scene. We demonstrate that in the depth modality, ViewSynth outperforms the state-of-the-art depth and RGB local feature extraction techniques in the 3D keypoint matching and camera localization tasks on the RGB-D datasets 7-Scenes, TUM RGBD and CoRBS in most scenarios. We also show the generalizability of ViewSynth in 3D keypoint matching across different datasets.

Marc Eder, True Price, Thanh Vu et al.

We present a versatile formulation of the convolution operation that we term a "mapped convolution." The standard convolution operation implicitly samples the pixel grid and computes a weighted sum. Our mapped convolution decouples these two components, freeing the operation from the confines of the image grid and allowing the kernel to process any type of structured data. As a test case, we demonstrate its use by applying it to dense inference on spherical data. We perform an in-depth study of existing spherical image convolution methods and propose an improved sampling method for equirectangular images. Then, we discuss the impact of data discretization when deriving a sampling function, highlighting drawbacks of the cube map representation for spherical data. Finally, we illustrate how mapped convolutions enable us to convolve directly on a mesh by projecting the spherical image onto a geodesic grid and training on the textured mesh. This method exceeds the state of the art for spherical depth estimation by nearly 17%. Our findings suggest that mapped convolutions can be instrumental in expanding the application scope of convolutional neural networks.

Marc Eder, Jan-Michael Frahm

Applying convolutional neural networks to spherical images requires particular considerations. We look to the millennia of work on cartographic map projections to provide the tools to define an optimal representation of spherical images for the convolution operation. We propose a representation for deep spherical image inference based on the icosahedral Snyder equal-area (ISEA) projection, a projection onto a geodesic grid, and show that it vastly exceeds the state-of-the-art for convolution on spherical images, improving semantic segmentation results by 12.6%.

Rui Wang, Stephen M. Pizer, Jan-Michael Frahm

Deep learning-based, single-view depth estimation methods have recently shown highly promising results. However, such methods ignore one of the most important features for determining depth in the human vision system, which is motion. We propose a learning-based, multi-view dense depth map and odometry estimation method that uses Recurrent Neural Networks (RNN) and trains utilizing multi-view image reprojection and forward-backward flow-consistency losses. Our model can be trained in a supervised or even unsupervised mode. It is designed for depth and visual odometry estimation from video where the input frames are temporally correlated. However, it also generalizes to single-view depth estimation. Our method produces superior results to the state-of-the-art approaches for single-view and multi-view learning-based depth estimation on the KITTI driving dataset.

Rui Wang, Jan-Michael Frahm, Stephen M. Pizer

Learning-based, single-view depth estimation often generalizes poorly to unseen datasets. While learning-based, two-frame depth estimation solves this problem to some extent by learning to match features across frames, it performs poorly at large depth where the uncertainty is high. There exists few learning-based, multi-view depth estimation methods. In this paper, we present a learning-based, multi-view dense depth map and ego-motion estimation method that uses Recurrent Neural Networks (RNN). Our model is designed for 3D reconstruction from video where the input frames are temporally correlated. It is generalizable to single- or two-view dense depth estimation. Compared to recent single- or two-view CNN-based depth estimation methods, our model leverages more views and achieves more accurate results, especially at large distances. Our method produces superior results to the state-of-the-art learning-based, single- or two-view depth estimation methods on both indoor and outdoor benchmark datasets. We also demonstrate that our method can even work on extremely difficult sequences, such as endoscopic video, where none of the assumptions (static scene, constant lighting, Lambertian reflection, etc.) from traditional 3D reconstruction methods hold.

Akash Bapat, Jan-Michael Frahm

We present a framework for edge-aware optimization that is an order of magnitude faster than the state of the art while having comparable performance. Our key insight is that the optimization can be formulated by leveraging properties of the domain transform, a method for edge-aware filtering that defines a distance-preserving 1D mapping of the input space. This enables our method to improve performance for a variety of problems including stereo, depth super-resolution, and render from defocus, while keeping the computational complexity linear in the number of pixels. Our method is highly parallelizable and adaptable, and it has demonstrable scalability with respect to image resolution.

Enliang Zheng, Dinghuang Ji, Enrique Dunn et al.

We target the problem of sparse 3D reconstruction of dynamic objects observed by multiple unsynchronized video cameras with unknown temporal overlap. To this end, we develop a framework to recover the unknown structure without sequencing information across video sequences. Our proposed compressed sensing framework poses the estimation of 3D structure as the problem of dictionary learning, where the dictionary is defined as an aggregation of the temporally varying 3D structures. Given the smooth motion of dynamic objects, we observe any element in the dictionary can be well approximated by a sparse linear combination of other elements in the same dictionary (i. e. self-expression). Moreover, the sparse coefficients describing a locally linear 3D structural interpolation reveal the local sequencing information. Our formulation optimizes a biconvex cost function that leverages a compressed sensing formulation and enforces both structural dependency coherence across video streams, as well as motion smoothness across estimates from common video sources. We further analyze the reconstructability of our approach under different capture scenarios, and its comparison and relation to existing methods. Experimental results on large amounts of synthetic data as well as real imagery demonstrate the effectiveness of our approach.