Dina Bashkirova

I am postdoctoral researcher in the Image and Video Computing Group at Boston University. I received my PhD in Computer Science at Boston University advised by Professor Kate Saenko. My research is focused on robust and semantically-consistent image generation and editing. I am also interested in label-efficient learning, as well as applications of Computer Vision for Good.

Prior to joining BU, I received my BSc and MSc in Computer Science at Kazan Federal University.

dbash [at] bu [dot] edu  /  CV  /  short CV  /  Google Scholar

My primary interests include generative models, image translation and label-efficient learning. My goal is to make generative models more expressive and controllable for artists of various skill levels. With this goal, I explore various ways to achieve fine-grained semantic control with label-efficient disentanglement and domain alignment. In addition to my work on generative models, I am also deeply passionate about the applications of AI for Environment.

We propose Lasagna, a layered image editing approach that allows controlled and language-guided object relighting. Lasagna achieves a controlled relighting via layered score distillation sampling that allows extracting diffusion model prior about lighting without changing other crucial aspects of the input image.

In this paper, we introduce MaskSketch, an image generation method that allows spatial conditioning of the generation result using a guiding sketch as an extra conditioning signal during sampling. Given an input sketch and its class label, MaskSketch samples realistic images that follow the given structure. MaskSketech works on sketches of various degrees of abstraction by leveraging a pre-trained masked image generators, while not requiring model finetuning or pairwise supervision.

We introduce a visual domain adaptation challenge for industrial waste sorting. Large variability in visual appearance and composition of waste stream make generalization especially challenging for the automated waste sorting solutions. In this competition, we propose to improve generalization of the waste detection models using SynthWaste, novel synthetic dataset for waste sorting.

We propose a new many-to-many image translation method that infers which attributes are domain-specific from data by constraining information flow through the network using translation honesty losses and a penalty on the capacity of the domain-specific embedding, and does not rely on hard-coded inductive architectural biases.

We present the first in-the-wild object segmentation dataset for industrial waste sorting for fully-, semi- and weakly-supervised setups. Our ZeroWaste datasets presents a challenging computer vision task of semantic segmentation of extremely cluttered scenes with highly deformable and translucent objects.

We propose an evaluation protocol for the disentanglement quality of unsupervised many-to-many image translation (UMI2I) methods. We show that modern UMI2I methods fail to correctly disentangle the domain-specific from shared factors and mostly rely on their corresponding inductive biases to determine which factors should be changed after translation.

The fifth iteration of the Visual Domain Adaptation (VisDA) Challenge in which we target the universal domain adaptation problem.

We propose a novel modular model architecture that allows parameter sharing and reuse for various computer vision tasks, such as multi-task learning, knowledge transfer and domain adaptation.

We show that cycle-consistent models perform a self-adversarial attack by embedding low-amplitude structured noise into intermediate generated images to reconstruct input images perfectly. We propose two techniques that prevent this kind of "cheating" and show that defending against such self-adversarial attacks improves the translation quality.

We propose a spatiotemporal extention of CycleGAN and show when it performs better then per-frame translation on two novel unsupervised video-to-video translation benchmarks including a novel CT-to-MRI volumetric medical domain.

We propose a novel approximation method for fast Gaussian convolution of two-dimensional uniform point sets that involves L1 distance metric for Gaussian function and domain splitting approach to achieve fast computation.

Grader for CS480 Introduction to Computer Graphics in 2018, CS542 Machine Learning in 2020 and 2021.

Reviewer for NeurIPS 2018-2021, ICCV 2021, ICLR 2020, 2021, 2023, CVPR 2020, 2023, WACV 2020, 2021, NeurIPS DistShift Workshop 2021. Co-organized the VisDA challenge at NeurIPS21 and NeurIPS22.

I enjoy reading, singing and playing music with my friends, drawing and handcrafting, weightlifting, playing boardgames and spending time with my dogs.

Template adapted from Jon Barron's homepage.