Deepak Kumar Singh

I'm a Computer Vision and Deep Learning Engineer at Mercedes-Benz Research and Development India, Bengaluru, part of the Interior-Sensing (IS) Algorithm team. I build production-grade computer vision systems for in-cabin automotive applications — including face recognition, depth estimation, and IR-based occupant analysis — owning the pipeline from proof-of-concept to deployment.

I completed my MS by Research at CVIT, IIIT Hyderabad, advised by Prof. C.V. Jawahar and co-advised by Prof. Vineeth N Balasubramanian, Prof. Chetan Arora, and Dr. Anbumani Subramanian. My current areas of interest are Face Recognition, Vision-Language Models, Contrastive Learning, Knowledge Distillation, and Open-World / Incremental Learning.

I like photography, majorly urban photography and landscapes; check out my Flickr. I've recently started learning Bansuri(Indian Flute). To know what I'm currently reading, checkout my Goodreads.

Email: deepaksinghcv@gmail.com

|    CV  /  LinkedIn  /  Github     | 

My current work spans in-cabin face recognition, cross-domain (IR↔RGB) learning, and occupant analysis for automotive systems. Prior research focused on Open-World Object Detection and benchmarking detection/segmentation models on road-scene datasets. I'm broadly interested in translating research into scalable, deployable systems. Authors marked with a * have an equal contribution in the research.

Member of the Interior-Sensing (IS) Algorithm team, building computer vision solutions for in-cabin automotive systems across face recognition, depth estimation, and weight prediction — owning the pipeline from proof-of-concept to production.

• Led end-to-end development of production-grade in-cabin face recognition systems, from data strategy and model development to deployment, achieving stringent accuracy and robustness targets.
• Architected a dual-level authentication framework enabling both low-latency personalization and high-assurance security for privacy- and payment-sensitive use cases.
• Developed cross-domain face recognition and IR-to-RGB colorization solutions to bridge RGB/NIR modality gaps and reduce dependency on large-scale RGB data collection.
• Built deep learning models for occupant analysis, including IR-based weight estimation, replacing hardware sensors with software-driven solutions.
• Collaborated across software, quantization, validation, and platform teams; drove GCP-based ML infrastructure initiatives and contributed to patentable innovations.

• Built the product's main JavaScript UI console and dynamic data-visualization modules (IgniteUI), and led migration of the in-house application from Adobe Flash to a JavaScript environment.
• Developed REST APIs in Java and data adapters powering live data-visualization modules.

Extends our open-world object detection work to incrementally adapt to new vehicle categories without catastrophic forgetting, improving robustness for real-world autonomous driving deployments.

This work formulates Open-World Object Detection by addressing the inherent issues present in road scene datasets India Driving Dataset(IDD) and Berkeley Deep Drive(BDD).

Benchmark latest models of object detection, semantic segmentation, and instance segmentation on road scene datasets. We perform a detailed study on their behaviour on both constrained and unconstrained datasets; Cityscapes, India Driving Dataset(IDD) and Berkeley Deep Drive(BDD).

• In-Cabin Face Authentication System — Filed in Germany and India
  Implemented Patent Award · Inventor lead. Enables secure multi-level authentication in automotive systems.


• Intelligent In-Vehicle Hyper-Personalization — Filed in Germany and India
  Co-Inventor. Enhances user experience by assimilating signals, memory, and preferences.

Built a fully-local multimodal RAG assistant for technical PDFs with two-stage retrieval (bi-encoder recall + cross-encoder reranking) and page-level citations, running entirely on-device with no external APIs. Figures are extracted, captioned by a local VLM, embedded into a unified index, and passed back to the vision model for image-grounded answers. A Recall@K / MRR evaluation harness over a 100-question gold set drove evidence-based tuning, reaching Recall@5 ~ 0.93 (reranked) and MRR ~ 0.74.

Fine-tuned Qwen3-VL-4B with QLoRA (4-bit, LoRA) to estimate calories and macros from food images, outputting structured JSON at 100% schema compliance on held-out data. A controlled experiment comparing two label sources cut calorie error to 81.4 kcal MAE (~34% lower than the base model) and reduced carbohydrate error by ~45% on a clean lab-measured holdout. Demonstrated that fine-tuning on synthetic class-average labels (Food101) degraded accuracy below the un-tuned baseline, while real measured data (Nutrition5k) produced genuine estimation — a data-quality finding with direct modeling implications.

• Kaizen (Operational Excellence) Recognition — Mar 2026
  Selected at department level for innovation leading to significant cost and efficiency gains.


• World Quality Week Recognition — Nov 2025
  Presented the benefits, optimizations, and achievements of our solutions.


• Customer Appreciation Recognition — Sep 2025
  Recognized by the Mercedes-Benz client team.


• Impact Award (Silver) — Feb 2025, 2023
  Recognized for delivering high-impact ML solutions.


• Team Excellence Award (Silver) — Oct 2024
  Awarded for collaborative delivery of production-grade systems.

Template from here and here. Thank you :)

Last Updated: 15^th June 2026