Aditya Deshpande

Learning Large-Scale Automatic Image Colorization

We describe an automated method for image colorization that learns to colorize from examples. Our method exploits a LEARCH framework to train a quadratic objective function in the chromaticity maps, comparable to a Gaussian random field. The coefficients of the objective function are conditioned on image features, using a random forest. The objective function admits correlations on long spatial scales, and can control spatial error in the colorization of the image. Images are then colorized by minimizing this objective function. We demonstrate that our method strongly outperforms a natural baseline on large-scale experiments with images of real scenes using a demanding loss function. We demonstrate that learning a model that is conditioned on scene produces improved results. We show how to incorporate a desired color histogram into the objective function, and that doing so can lead to further improvements in results.

Learning Diverse Image Colorization

Colorization is an ambiguous problem, with multiple viable colorizations for a single grey-level image. However, previous methods only produce the single most probable colorization. Our goal is to model the diversity intrinsic to the problem of colorization and produce multiple colorizations that display long-scale spatial co-ordination. We learn a low dimensional embedding of color fields using a variational autoencoder (VAE). We construct loss terms for the VAE decoder that avoid blurry outputs and take into account the uneven distribution of pixel colors. Finally, we build a conditional model for the multi-modal distribution between grey-level image and the color field embeddings. Samples from this conditional model result in diverse colorization. We demonstrate that our method obtains better diverse colorizations than a standard conditional variational autoencoder (CVAE) model, as well as a recently proposed conditional generative adversarial network (cGAN).

Multistage SFM: Revisiting Incremental Structure from Motion

In this paper, we present a new multistage approach for SfM reconstruction of a single component. Our method begins with building a coarse 3D reconstruction using high-scale features of given images. This step uses only a fraction of features and is fast. We enrich the model in stages by localizing remaining images to it and matching and triangulating remaining features. Unlike traditional incremental SfM, localization and triangulation steps in our approach are made efficient and embarrassingly parallel using geometry of the coarse model. The coarse model allows us to use 3D-2D correspondences based direct localization techniques to register remaining images. We further utilize the geometry of the coarse model to reduce the pair-wise image matching effort as well as to perform fast guided feature matching for majority of features. Our method produces similar quality models as compared to incremental SfM methods while being notably fast and parallel. Our algorithm can reconstruct a 1000 images dataset in 15 hours using a single core, in about 2 hours using 8 cores and in a few minutes by utilizing full parallelism of about 200 cores.

Can GPUs sort strings efficiently

String sorting or variable-length key sorting has lagged in performance on the GPU even as the fixed-length key sorting has improved dramatically. Radix sorting is the fastest on the GPUs. In this paper, we present a fast and efficient string sort on the GPU that is built on the available radix sort. Our method sorts strings from left to right in steps, moving only indexes and small prefixes for efficiency. We reduce the number of sort steps by adaptively consuming maximum string bytes based on the number of segments in each step. Performance is improved by using Thrust primitives for most steps and by removing singleton segments from consideration. Over 70% of the string sort time is spent on Thrust primitives. This provides high performance along with high adaptability to future GPUs. We achieve speed of up to 10 over current GPU methods, especially on large datasets. We also scale to much larger input sizes. We present results on easy and difficult strings defined using their after-sort tie lengths.

Hybrid implementation of error diffusion dithering

Many image filtering operations provide ample parallelism, but progressive non-linear processing of images is among the hardest to parallelize due to long, sequential, and non-linear data dependency. A typical example of such an operation is error diffusion dithering, exemplified by the Floyd-Steinberg algorithm. In this paper, we present its parallelization on multicore CPUs using a block-based approach and on the GPU using a pixel based approach. We also present a hybrid approach in which the CPU and the GPU operate in parallel during the computation. High Performance Computing has traditionally been associated with high end CPUs and GPUs. Our focus is on everyday computers such as laptops and desktops, where significant compute power is available on the GPU as on the CPU. Our implementation can dither an 8K × 8K image on an off-the-shelf laptop with an Nvidia 8600M GPU in about 400 milliseconds when the sequential implementation on its CPU took about 4 seconds.

Geometry directed browser for personal photographs

Browsers of personal digital photographs all essentially follow the slide show paradigm, sequencing through the photos in the order they are taken. A more engaging way to browse personal photographs, especially of a large space like a popular monument, should involve the geometric context of the space. In this paper, we present a geometry directed photo browser that enables users to browse their personal pictures with the underlying geometry of the space to guide the process. The browser uses a pre-computed package of geometric information about the monument for this task. The package is used to register a set of photographs taken by the user with the common geometric space of the monument. This involves localizing the images to the monument space by computing the camera matrix corresponding to it. We use a state-of-the-art method for fast localization. Registered photographs can be browsed using a visualization module that shows them in the proper geometric context with respect to a point-based 3D model of the monument. We present the design of the geometry-directed browser and demonstrate its utility for a few sequences of personal images of well-known monuments. We believe personal photo browsers can provide an enhanced sense of ones own experience with a monument using the underlying geometric context of the monument.

Fast Burrows Wheeler Compression Using All-Cores

In this paper, we present an all-core implementation of Burrows Wheeler Compression algorithm that exploits all computing resources on a system. Our focus is to provide significant benefit to everyday users on common end-to-end applications by exploiting the parallelism of multiple CPU cores and additional accelerators, viz. Many-core GPU, on their machines. The all-core framework is suitable for problems that process large files or buffers in blocks. We consider a system to be made up of compute stations and use a work-queue to dynamically divide the tasks among them. Each compute station uses an implementation that optimally exploits its architecture. We develop a fast GPU BWC algorithm by extending the state-of-the-art GPU string sort to efficiently perform BWT step of BWC. Our hybrid BWC with GPU acceleration achieves a 2.9× speedup over best CPU implementation. Our all-core framework allows concurrent processing of blocks by both GPU and all available CPU cores. We achieve a 3.06× speedup by using all CPU cores and a 4.87× speedup when we additionally use an accelerator i.e. GPU. Our approach will scale to the number and different types of computing resources or accelerators found on a system.

Recovering the 3D Geometry of Heritage Monuments from Image Collections

Several methods have been proposed for large-scale 3D reconstruction from large, unorganized image collections. A large reconstruction problem is typically divided into multiple components which are reconstructed independently using structure from motion (SFM) and later merged together. Incremental SFM methods are most popular for the basic structure recovery of a single component. They are robust and effective but strictly sequential in nature. We present a multistage approach for SFM reconstruction of a single component that breaks the sequential nature of the incremental SFM methods. Our approach begins with quickly building a coarse 3D model using only a fraction of features from given images. The coarse model is then enriched by localizing remaining images and matching and triangulating remaining features in subsequent stages. The geometric information available in the form of the coarse model allows us to make these stages effective, efficient, and highly parallel. We show that our method produces similar quality models as compared to standard SFM methods while being notably fast and parallel.

Top Down Approach to Detect Multiple Planes from Pair of Images

Detecting multiple planes in images is a challenging problem, but one with many applications. Recent work such as J-Linkage and Ordered Residual Kernels have focussed on developing a domain independent approach to detect multiple structures. These multiple structure detection methods are then used for estimating multiple homographies given feature matches between two images. Features participating in the multiple homographies detected, provide us the multiple scene planes. We show that these methods provide locally optimal results and fail to merge detected planar patches to the true scene planes. These methods use only residues obtained on applying homography of one plane to another as cue for merging. In this paper, we develop additional cues such as local consistency of planes, local normals, texture etc. to perform better classification and merging. We formulate the classification as an MRF problem and use TRWS message passing algorithm to solve non metric energy terms and complex sparse graph structure. We show results on Michigan Indoor Corridor Dataset and our challenging dataset, common in robotics navigation scenarios. Experiments on the datasets demonstrate the accuracy of our plane detection relative to ground truth, with detailed comparisons to prior art.