Research Overview
My Research
Geometry is easier to capture than to use. A mesh can reproduce the surface of a chair, yet say nothing about which legs should remain symmetric, which parts must stay attached, or which dimensions a designer should be able to change. For editing, fabrication, and reasoning, these relationships matter as much as the visible shape.
My research lies at the intersection of computer vision, computer graphics, and machine learning. I work on geometry we can act on: representations that connect visual observations to explicit parts, parameters, constraints, relationships, and executable operations. I study these as symbolic geometry—models that complement images, meshes, and point clouds by exposing meaningful handles for changing, checking, and reusing geometry.
PhD Thesis
My PhD thesis, From Measurement to Structure: Annotation-Free Acquisition of Symbolic Geometry, focuses on a central bottleneck: acquisition. New symbolic representations are often introduced precisely because existing formats omit the structure a task needs. But a new language often arrives before its corpus. I study annotation-free acquisition: recovering, constructing, or fitting symbolic structure without relying on a large dataset that pairs inputs with hand-authored descriptions in the target representation.
The thesis develops three recurring sources of acquisition signal:
In-domain synthetic supervision. The representation itself can become a teacher. SIRI uses code rewriting to turn imperfect inferred programs into better training examples, while Pattern Analogies samples programmatic before-and-after edits and learns to transfer their effects to real patterns whose original programs are unavailable.
Foundation-model priors. Pretrained models can provide semantic hypotheses, but plausible suggestions are not yet reliable geometry. ParSEL uses a language model to interpret a natural-language 3D edit, then grounds and completes the proposal with part relations and geometric reasoning to produce a coherent parameterized editing program.
Direct optimization. The target instance and its constraints can provide the signal. MiGumi co-designs a milling-aware language and optimization objectives to convert idealized joint programs into CNC-millable, tightly coupled programs. SuperFit co-designs an expressive primitive family and fitting procedure to recover compact, editable assemblies from meshes.
The broader lesson. Representations should be designed not only to express useful structure, but also to make that structure recoverable. When a language can be executed, rewritten, sampled, constrained, or optimized, it gives learning and search a foothold—and lets us build geometry that is not merely accurate to look at, but useful to edit, fabricate, and act on.
Past Research
Previously, I worked as a Researcher at Preferred Networks, Japan, and, before that, as a Research Assistant at Video Analytics Lab, IISc, Bangalore, India. I have worked on various projects in Computer Vision, Machine Learning, Audio Processing, Human-Computer Interaction, and Robotics. I have been really fortunate to collaborate and get mentorship from some amazing people - Daniel Ritchie, R Venkatesh Babu, Siddhartha Chaudhuri, Shin-ichi Maeda, Adrien Gaidon, Rares Ambrus, Jason Naradowsky, and Fabrice Matulic to name a few.
Some of the central themes in my previous research:
- Novel methods for Computer Vision tasks (Semantic Segmentation, 3D Object Pose estimation, etc.),
- Stability of deep learnt features (Adversarial Attacks),
- Application of DL in Audio (source separation) and HCI (Tracking in VR).
For a complete list of published papers, visit Google Scholar.
Publications:
Residual Primitive Fitting of 3D Shapes with SuperFrusta
A. Ganeshan, Matheus Gadelha, Thibault Groueix, Zhiqin Chen, Siddhartha Chaudhuri, Vladimir G. Kim, Wang Yifan and Daniel Ritchie
Oral & Award Candidate (1.8%) IEEE / CVF Computer Vision and Pattern Recognition Conference (CVPR) 2026
We introduce a framework for converting 3D shapes into compact and editable assemblies of analytic primitives. Our approach combines SuperFrustum, an expressive, editable, and optimizable primitive, with ResFit, an iterative fitting algorithm that achieves state-of-the-art reconstruction quality while using significantly fewer primitives than prior work.
MiGumi - Making Tightly Coupled Integral Joints Millable
A. Ganeshan, Kurt Fleischer, Wenzel Jakob, Ariel Shamir, Daniel Ritchie, Takeo Igarashi and Maria Larsson
ACM Siggraph Asia 2025, Journal at Transactions on Graphics (TOG) 2025
We introduce MXG, a language for representing geometry as milling operations with flat-end drill bits, enabling traditional integral wood joints to be adapted for CNC fabrication. Through gradient-based co-optimization of part geometries, we restore tight coupling between parts despite milling-induced artifacts like rounded corners.
Procedural Scene Programs for Open-Universe Scene Generation - LLM-Free Error Correction via Program Search
Maxim Gumin, Do Heon Han, Seung Jean Yoo, A. Ganeshan, R. Kenny Jones, Kailiang Fu, Rio Aguina-Kang, Stewart Morris, Daniel Ritchie
ACM Siggraph Asia 2025, Journal at Transactions on Graphics (TOG) 2025
We introduce procedural scene programs, an imperative paradigm where LLMs iteratively place objects as functions of previously-placed objects, handling wider variety and complexity than declarative constraint-based approaches. Our LLM-free error correction via program search iteratively improves scene validity while preserving layout intent, with participants preferring our approach 82-94% of the time.
Pattern Analogies - Learning to Perform Programmatic Image Edits by Analogy
A. Ganeshan, Thibault Groueix, Paul Guerrero, Radomír Měch, Matthew Fisher and Daniel Ritchie
IEEE / CVF Computer Vision and Pattern Recognition Conference (CVPR) 2025
We enable programmatic editing of pattern images through pattern analogies—pairs of simple patterns demonstrating the intended edit. Using our SplitWeave DSL and TriFuser (a specialized Latent Diffusion Model), our method performs structure-aware edits while generalizing to unseen pattern styles.
ParSEL - Parameterized Shape Editing with Language
A. Ganeshan, Ryan Y. Huang, Xianghao Xu, R. Kenny Jones and Daniel Ritchie
ACM Siggraph Asia 2024, Journal at Transactions on Graphics (TOG) 2024.
We enable controllable editing of 3D shapes from natural language by producing parameterized editing programs users can adjust to explore variations. Our Analytical Edit Propagation (AEP) algorithm leverages computer algebra systems to extend LLM-generated seed edits into complete, semantically-valid programs.
Learning to Edit Visual Programs with Self-Supervision
R Kenny Jones, Renhao Zhang, A. Ganeshan, Daniel Ritchie
Conference on Neural Information Processing Systems (NeurIPS), 2024
We train an edit network that predicts local edit operations to improve visual programs' similarity to target images, learning through self-supervised bootstrapped finetuning. Our editing-based inference paradigm, which evolves populations initialized from one-shot predictions, significantly outperforms using only one-shot models.
Creating Language-driven Spatial Variations of Icon Images
Xianghao Xu, A. Ganeshan, Karl DD Willis, Yewen Pu, Daniel Ritchie
Arxiv
We enable language-driven spatial editing of icon images by translating user requests into geometric constraints using LLMs and a DSL library. Optimizing affine transformations of icon segments with respect to these constraints produces spatially varied icons while preserving physical and logical coherence.
Open-universe indoor scene generation using llm program synthesis and uncurated object databases
Rio Aguina-Kang, Maxim Gumin, Do Heon Han, Stewart Morris, Seung Jean Yoo, A. Ganeshan, R Kenny Jones, Qiuhong Anna Wei, Kailiang Fu, Daniel Ritchie
Arxiv
We generate indoor scenes from text prompts by leveraging LLMs to synthesize programs in a layout DSL, solving constraints via gradient optimization, and retrieving meshes from uncurated databases using VLMs. Our open-universe approach handles arbitrary prompts and object categories without requiring large 3D scene training datasets.
Improving Unsupervised Visual Program Inference with Code Rewriting Families
A. Ganeshan, R. Kenny Jones and Daniel Ritchie
Oral (9.03%) IEEE / CVF International Conference on Computer Vision (ICCV), 2023
We introduce SIRI, a framework for bootstrapped learning that improves visual program inference by sparsely applying code rewriters—Parameter Optimization, Code Pruning, and Code Grafting. Our family of rewriters produces more parsimonious programs with fewer primitives while matching or outperforming domain-specific neural architectures.
Skill Generalization with Verbs
R. Ma, L. Lam, B. A. Spiegel, A. Ganeshan, B. Abbatematteo, R. Patel, D. Paulius, S. Tellex, G. Konidaris.
IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2023
We enable robots to generalize manipulation skills to novel objects using verbs as semantic abstractions. Our probabilistic classifier learns verb-trajectory associations and performs policy search to generate executable trajectories for new object instances.
Warp-Refine Propagation Semi-Supervised Auto-labeling via Cycle-consistency
A. Ganeshan , Alexis Vallet, Yasunori Kudo, Shin-ichi Maeda, Tommi Kerola, Rares Ambrus, Dennis Park, Adrien Gaidon
IEEE / CVF International Conference on Computer Vision (ICCV), 2021
We propose Warp-Refine Propagation, a label propagation method that combines geometric warping with learned semantic priors via cycle consistency for semi-supervised video annotation. Our approach improves label propagation by 13.1 mIoU on ApolloScape and achieves competitive results on semantic segmentation benchmarks.
Phonetroller - Visual Representations of Fingers for Precise Touch Input with Mobile Phones in VR
Fabrice Matulic, A. Ganeshan, Hiroshi Fujiwara, Daniel Vogel
Conference on Human Factors in Computing Systems, CHI 21
We enable precise touch input on smartphones in fully immersive VR by mounting a mirror above the screen to capture thumbs via the front camera. Deep learning infers fingertip positions, providing semi-transparent overlays that guide users to thumb-size targets without seeing their hands.
Meta-Learning Extractors for Music Source Separation
David Samuel, A. Ganeshan, Jason Naradowsky
IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP-2020)
We propose Meta-TasNet, a hierarchical meta-learning model where a generator predicts weights for instrument-specific extractors in music source separation. This enables efficient parameter-sharing while achieving performance comparable to state-of-the-art with fewer parameters and faster runtime.
FDA Feature Disruptive Attack
A. Ganeshan, B. S. Vivek, R. V. Babu
IEEE / CVF International Conference on Computer Vision (ICCV), 2019
We propose Feature Disruptive Attack (FDA), an adversarial attack that corrupts features at each network layer rather than targeting outputs. FDA generates stronger adversaries that severely degrade deep network performance and disrupt feature-representation tasks even without access to task-specific networks.
Enhancing Semantic Segmentation by Learning Expertise between Confusing Classes
A. Ganeshan, G. S. Rajput, R. V. Babu
First International Workshop On Autonomous Navigation in Unconstrained Environments, ECCV 18
We propose an Expertise-Layer to enhance semantic segmentation models' ability to discern differences between confusing classes in unconstrained datasets with high intra-class variations. Our approach improves performance on real-world scenarios with similar classes and low-shot categories.
Object Pose Estimation from Monocular Image using Multi-View Keypoint Correspondence
J. N. Kundu*, A. Ganeshan*, M. V. Rahul*, R. V. Babu
Geometry Meets Deep Learning Workshop, at ECCV 2018.
We learn pose-invariant local descriptors from 2D RGB images and generate keypoint correspondence maps using renders of 3D template models. Fusing multi-view correspondence information improves geometric comprehension and enables state-of-the-art pose estimation while alleviating data scarcity.
Generalizable Data-free Objective for Crafting Universal Adversarial Perturbations
K. M. Reddy*, A. Ganeshan*, R. V. Babu
IEEE Transactions on Patter Analysis and Machine Intelligence, 2018
We present a generalizable, data-free objective for crafting universal adversarial perturbations that works across multiple vision tasks by corrupting features at multiple layers. Our method outperforms data-dependent approaches in black-box scenarios and achieves significant fooling rates without requiring training data.
iSPA-Net Iterative Semantic Pose Alignment Network
J. N. Kundu*, A. Ganeshan*, M. V. Rahul*, R. V. Babu
ACM International Conference on Multimedia 2018
We estimate fine-grained 3D object pose by predicting viewpoint differences between image pairs, leveraging semantic structural regularity from 3D CAD models. Our iterative refinement framework uses online rendering and non-uniform bin classification to achieve state-of-the-art performance on real image datasets.