Prabhat Kumar

Leading research and engineering in multi-modal video understanding architectures.

• Researched multi-modal Scene Break Detection, scaling temporal sequence modeling by benchmarking linear-complexity State-Space Models (Mamba) and GCNs against standard Transformers.
• Formulated self-supervised proxy tasks to align asynchronous cross-modal representations from visual (DINO), textual (Sentence Transformers), and acoustic foundation models for precise boundary localization.
• Scaled distributed feature extraction and experimental validation across ultra-long-form video datasets using Ray, enabling high-throughput parallel processing and localized structural reasoning.

Promoted to Tech Lead (Jul 2025); served as R&D Specialist (Oct 2022 – Jun 2025) driving autonomous perception.

• Developed unsupervised cross-camera optical flow estimation by coupling a UNIT-GAN translation network with a FlowFormer backbone, enabling zero-shot model and label transfer across heterogeneous sensors.
• Formulated dual-level domain adaptation objectives using predicted dense flow fields to warp visual representations at both pixel and latent feature levels, maximizing cross-domain geometric consistency.
• Leveraged flow-based feature alignment to jointly regularize and improve representation capacity across source and target camera perception networks, resulting in 3 published patents.

Developed onboard computer vision and multi-task learning for Autonomous Driving for Indian Roads.

• Developed a unified multi-task network for joint Semantic Segmentation and Depth Estimation tailored for unstructured road environments using hybrid supervised and self-supervised training.
• Implemented domain adaptation pipelines and novel geometric and photometric data augmentations to minimize feature degradation against severe out-of-distribution driving conditions.
• Evaluated model efficiency and real-world inference constraints by adapting neural architectures for low-latency execution via TensorRT within an onboard ROS robotics environment.

Foundational academic research in self-supervised representation learning and deepfake forensics.

• Indian Institute of Science (IISc): Formulated a self-supervised representation learning framework (SISL) to decouple content-independent camera signatures from image patches, training an anomaly localization network to identify splicing boundaries without pixel-level ground truth.
• IIT Jodhpur: Developed a multi-stream deep neural network to isolate regional artifacts and spatial-temporal texture distortions from facial reenactment pipelines, mapping model robustness across progressive video compression codecs (published in WACV 2020).