Anidh Singh

Collaborated with the team to implement various super-resolution research papers to establish a baseline for superresolution models.• Researched and analysed various academic papers to arrive at a QoE metric which could help us quantify theperceptual quality of the output video from the trained models.• Lead the effort to design GPU model architectures for different upscaling factors using tensorflow with theconstraint of making them run in real-time on various mid-tier non-qualcomm devices.• Applied the quantization aware training code to Efficient Sub-pixel Convolution Neural Network (ESPCN) modelthereby avoiding the quantization degradation on the edge run time environments like Digital SignalProcessor(DSP) by 15% and enabling the models to run faster.• Spearheaded the effort of doing super resolution using android specific renderscript library. This enabled us to dosuper resolution on extremely low compute edge devices and nearly match the performance of the tensorflow litemodels.• Strategized models for 4K content generation from low bitrate videos. Implemented the whole pipeline in Nvidiaagx xavier board leveraging the tensorrt framework.• Created new datasets for super resolution models to overcome the deficiencies of previous datasets by leveraging theTfrecords API and thereby increasing the overall perceptual quality by about 10 percent.• Developed a docker container and scripts to do inference on edge devices which reduced the team’s time and effortby 80%.