Mohit Goyal

We introduce SPLIT, a training-free harness that reconstructs one metric 3-D point cloud from a video and gives a planner VLM tools to query it. Frozen reconstruction and grounding models locate objects and measure their geometry, and the planner VLM combines those measurements using Python. SPLIT improves accuracy on VSIBench, VSTIBench, and ReVSI, achieving the best published training-free score on VSIBench. On VSTIBench, it raises accuracy from 57.7% to 78.3%, eliminating 49% of the baseline VLM's errors. Rerunning SPLIT with ground-truth measurements in place of the tools' estimates shows that inaccurate perception, not reasoning, remains a major source of error.

@misc{yeung2026split,
  title = {Perception, Not Reasoning, Limits Video Spatial Understanding},
  author = {Yeung, Jacob and Goyal, Mohit and Dubost, Florian and Vondran Jr., Gary L. and Tombari, Federico and Ramanan, Deva and Tarr, Michael J.},
  year = {2026}
}

MLLMs struggle with fine-grained spatial understanding such as route tracing on maps, in part due to the cost of collecting pixel-accurate path annotations. We introduce a scalable synthetic data generation pipeline that leverages synthetic map images, pixel-graph shortest-path search, and MLLM-as-a-Judge critics to automatically produce precise annotations. Fine-tuning open-source (Gemma 3 27B) and proprietary (Gemini 2.5 Flash) MLLMs on the resulting data raises success rates on MapBench by up to 6.4 points while reducing path-tracing error (NDTW) by 32.5%.

This paper investigates the impact of viewing the world through video passthrough systems on human behavioral and physiological patterns through a large-scale multimodal study (N=110). We capture synchronized oculomotor, kinematic, and physiological data during a block assembly task. Our results reveal a four-fold suppression of rotational head velocity and a pronounced decoupling of head-gaze coordination, which we term "Passthrough Rigidity", providing a quantitative framework for evaluating and improving future XR devices.

Mind the GAP is the first work that demonstrates that GAP significantly reduces nausea, disorientation, and total scores of cybersickness as well as subjective discomfort scores as compared to DP. We present a comprehensive protocol aimed at evaluating visually-induced discomfort and cybersickness in VST HMDs through key use cases.

Interactive object understanding, or what we can do to objects and how is a long-standing goal of computer vision. In this paper, we tackle this problem through observation of human hands in in-the-wild egocentric videos. We apply these basic principles on the EPIC-KITCHENS dataset, and successfully learn state-sensitive features, and object affordances (regions of interaction and afforded grasps), purely by observing hands in egocentric videos.

@inproceedings{goyal2022human,
  author = "Goyal, Mohit and Modi, Sahil and Goyal, Rishabh and Gupta, Saurabh",
  title = "Human Hands as Probes for Interactive Object Understanding",
  year = "2022",
  booktitle = "Computer Vision and Pattern Recognition (CVPR)"
}

We propose JIND, a framework for automated cell-type identification based on neural networks that directly learns a low-dimensional representation in which cell-types can be reliably determined, using asymmetric adversarial batch alignment and calibrated rejection thresholds. In the reported experiments, JIND runs 2 to 6 times faster than ItClust.

DZip is a general-purpose lossless compressor for sequential data combining semi-adaptive biGRU + residual probability estimators with arithmetic coding, achieving 29% average size reduction over Gzip and 33% over LSTM-Compress across evaluated real datasets, without requiring external training data.

@INPROCEEDINGS{9418692,
  author={Goyal, Mohit and Tatwawadi, Kedar and Chandak, Shubham and Ochoa, Idoia},
  booktitle={2021 Data Compression Conference (DCC)},
  title={DZip: improved general-purpose lossless compression based on novel neural network modeling},
  year={2021},
  pages={153-162},
  doi={10.1109/DCC50243.2021.00023}}

We combine recurrent neural network (biGRU/LSTM) conditional probability estimators with an arithmetic coder to losslessly compress synthetic, text, and genomic sequences, outperforming Gzip and achieving near-optimal compression on synthetic benchmarks.

@inproceedings{goyal2019deepzip,
  title = "DeepZip: Lossless Data Compression Using Recurrent Neural Networks",
  author = "Mohit Goyal and Kedar Tatwawadi and Shubham Chandak and Idoia Ochoa",
  year = "2019",
  booktitle = "Proceedings - DCC 2019"
}

Formulates Glottal Closure Instant (GCI) detection directly from raw speech waveforms as a supervised multi-task learning problem using dilated convolutional neural networks jointly optimizing classification and regression objectives.