Srikumar Sastry

I am a PhD candidate at Washington University in Imaging Science, working in Multimodal Vision Research Laboratory led by Dr. Nathan Jacobs.

I have an MS in Geoinformatics from The Faculty ITC, Geoinformation Science and Earth Observation. During my Masters, I was supervised by Dr. Mariana Belgiu and Dr. Raian Vargas Maretto, and worked on developing active learning methods in remote sensing.

Email  /  CV  /  Scholar  /  Github  /  Academic Service

News

• Sep 2026: Heading to Amazon, Seattle as Applied Scientist Intern to work on multimodal models for perception.
• May 2026: Received Silver Reviewer Award at ICML 2026.
• May 2026: Heading to Zendar as PhD ML Intern to work on multimodal world models.
• Apr 2026: Accepted to Doctoral Consortium at CVPR 2026.
• Nov 2025: Applied for a patent for the work done at Dolby Laboratories titled "Probabilistic Multimodal Learning with Bayesian Disagreements".
• Sep 2025: Received travel grant to attend ICCV 2025.
• May 2025: Started as a PhD research intern at Dolby Laboratories, Sunnyvale.
• May 2025: Awarded as outstanding researcher at Imaging Science PhD awards 2025.
• Nov 2024: Received Top Reviewer Award at NeurIPS 2024.

Research

I am interested in multimodal learning, computer vision, remote sensing and generative/bayesian modeling. My work involves multimodal representation learning and fusing heterogeneous multimodal data for remote sensing applications involving fine-grained classification, satellite image synthesis, species distribution modeling, change detection, and land cover mapping.

We propose TerraDiT-Ω, a unified spatial control framework for satellite image generation that directly leverages native geospatial primitives—including points, bounding boxes, polylines, and polygons—enabling controllable synthesis across varying annotation budgets while remaining compatible with GeoAI workflows. TerraDiT-Ω introduces Geometry-Aware Local Attention to inject explicit geometric cues into the generation process, achieving state-of-the-art performance across conditioning formats and providing a versatile foundation for controllable data augmentation in diverse remote sensing tasks.

We propose TerraDiT, a diffusion transformer for text-to-satellite image generation with point-based control, enabling semantically rich and spatially precise generation using only point locations and textual descriptions instead of dense pixel-level maps. TerraDiT introduces an adaptive local attention mechanism to incorporate point queries effectively, achieving state-of-the-art performance while providing a flexible, annotation-efficient, and computationally simple framework for controllable satellite image synthesis.

We introduce ProM3E, a probabilistic masked multimodal embedding model that enables any-to-any modality generation, fusion feasibility analysis, and superior cross-modal retrieval and representation learning through masked modality reconstruction in the embedding space.

We propose SimLBR, a simple and efficient fake image detection framework that learns a tight decision boundary around the real image distribution and treats fake images as a sink class, improving robustness under distribution shift.

We propose PMBD, a probabilistic multimodal learning framework that addresses multiplicity, misalignment, and noise in multimodal data using a novel Bayesian Disagreement loss and Product of Experts fusion. It also introduces an efficient attention reuse algorithm and a Probabilistic Attention mechanism (PMBD-PA) for fine-grained token-level uncertainty quantification.

We propose a hierarchical multimodal contrastive learning framework that avoids false negatives across taxonomic levels by restricting comparisons within the same level. Combined with a group-balanced optimization strategy, the method substantially improves hierarchical consistency and zero-shot classification accuracy from coarse to fine granularity.

A state-of-the-art soundscape mapping framework that leverages a Vision-Language Model (VLM) to enrich the semantic understanding of a location’s soundscape, learns a shared codebook for fine-grained alignment, and enables retrieval-based, location-conditioned soundscape generation.

We consider the task of active geo-localization (AGL) in which an agent uses a sequence of visual cues observed during aerial navigation to find a target specified through multiple possible modalities.

We introduce VectorSynth, a diffusion-based model that generates pixel-accurate satellite imagery from polygonal geographic annotations with semantic attributes, enabling spatially grounded, geometry- and language-guided image synthesis and editing with superior semantic and structural fidelity.

In this work, we introduce Radial Cross-Modal Embeddings (RCME), a framework that enables the explicit modeling of transitivity-enforced entailment. Our proposed framework optimizes for the partial order of concepts within vision-language models.

We propose a novel retrieval-augmented strategy called RANGE, to fully capture the important visual features necessary to represent a geographic location.

TaxaBind is a suite of multimodal models useful for downstream ecological tasks covering six modalities: ground-level image, geographic location, satellite image, text, audio, and environmental features.

We consider the task of active geo-localization (AGL) in which an agent uses a sequence of visual cues observed during aerial navigation to find a target specified through multiple possible modalities.

A soundscape is defined by the acoustic environment a person perceives at a location. In this work, we propose a framework for mapping soundscapes across the Earth.

In this work, we present a deep-learning model, GeoBind, that can infer about multiple modalities, specifically text, image, and audio, from satellite imagery of a location. To do this, we use satellite images as the binding element and contrastively align all other modalities to the satellite image data.

We present GeoSynth, a model for synthesizing satellite images with global style and image-driven layout control. The global style control is via textual prompts or geographic location. These enable the specification of scene semantics or regional appearance respectively, and can be used together.

We propose a weakly supervised approach for creating maps using free-form textual descriptions. We refer to this work of creating textual maps as zero-shot mapping.

In this work, we present a model that can synthesize the visual appearance of the Earth's surface conditioned on climate and time. We use PRISM and temporally rich cloud-free Sentinel-2 Level 2A and LandSat-8 for this study.

We propose a novel model called Vision-Language Pseudo-Labeling (VLPL) which uses a vision-language model to suggest strong positive and negative pseudo-labels.

We focus on the problem of species distribution modeling using global-scale presence-only data. To capture a stronger implicit relationship between species, we encode the taxonomic hierarchy of species using a large language model.

We propose a metadata-aware self-supervised learning (SSL) framework useful for fine-grained classification and ecological mapping of bird species around the world. Our framework unifies two SSL strategies: Contrastive Learning (CL) and Masked Image Modeling (MIM), while also enriching the embedding space with metadata available with ground-level imagery of birds.

We focus on the task of soundscape mapping, which involves predicting the most probable sounds that could be perceived at a particular geographic location. We utilise recent state-of-the-art models to encode geotagged audio, a textual description of the audio, and an overhead image of its capture location using contrastive pre-training.

We consider the problem of cost effective active learning for semantic segmentation, which aims at reducing the efforts of semantically annotating images.

Miscellanea

Service

Teaching

Awesome Huggingface Demos