Varsha Hedau

3Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Varsha Hedau focuses on computer vision, specifically the 3D reconstruction and spatial understanding of indoor scenes from monocular images. Her research trajectory explores methods for recovering room layouts, modeling geometric context, and detecting objects within cluttered environments. The most recent work in the evidence addresses the recovery of free space by localizing furniture and floor contact points in 3D.

Research Trajectory

Hedau’s research addresses the complex challenge of extracting three-dimensional spatial information from single, two-dimensional images of indoor environments. A recurring obstacle in this domain is visual clutter, which frequently occludes structural boundaries such as the intersection of walls and floors.

In 2009, Hedau led work on this foundational problem as the first author of Recovering the spatial layout of cluttered rooms. To gain robustness against occlusion, this research modeled the global room space using a parametric 3D "box." Rather than relying solely on visible ground-wall boundaries, the approach utilized a structured learning algorithm that leveraged global perspective cues to iteratively localize clutter and fit the box parameters, thereby estimating the overall room geometry.

Building upon this geometric understanding of room layouts, Hedau’s subsequent work explored how global scene structure could inform the detection of specific objects within the room. In the first-author paper Thinking Inside the Box: Using Appearance Models and Context Based on Room Geometry (2010), she demonstrated that combining 3D perspective cues derived from scene geometry with traditional image-based detectors yields notably improved object detection. This work introduced a probabilistic model constrained by the spatial layout—specifically the estimated locations of walls and floors—to refine 3D object estimates, unifying contextual and geometric information.

By 2012, Hedau’s focus expanded from identifying objects and room boundaries to understanding the functional, navigable areas of an environment. As first author of Recovering free space of indoor scenes from a single image, she tackled the problem of delineating free space. This research introduced a "boxy" detector designed to localize box-shaped furniture oriented parallel to the scene across various scales. By identifying and blocking out the space occupied by these objects, the system could infer the remaining free space. A key technical contribution of this work was the design of specialized features that capture the floor contact points of objects, allowing for more accurate 3D localization than image-based metrics alone could provide.

Current Directions

The most recent paper in the provided evidence is from 2012. Based on this work, Hedau’s research investigates the functional understanding of indoor environments, specifically the delineation of free space versus occupied space. By focusing on the geometric structure of furniture and their physical grounding in the scene—such as floor contact points—this direction emphasizes the extraction of actionable 3D spatial information from single monocular images, moving beyond simple bounding boxes to a more holistic understanding of physical room capacity and layout.

Profile updated

See something wrong? Report a mistake