Apr 2021 – Mar 2024 · 2 yrs 11 mos

Apr 2017 – Apr 2021 · 4 yrs

• HEVC Encoder
• RDOQ for quantization coeff selection → RDOQ (Rate Distortion Optimised Quantization)
• chooses quantization coeff levels by choosing the local minima of the RDO cost function
• in the relevant parameter space. The challenge here was to invent an algorithm that
• minimises the points searched in the parameter space whilst also minimising the ’quality’
• loss relative to the brute force algorithm.
• Mode selection algorithms → Every CU in HEVC can be coded via a large variety of
• modes. Multiple sets of dovetailing algorithms, operating via different degrees of ’noisy’
• cost functions, were developed and used for selecting a ’good’ mode. Some of the algorithms
• were heuristics-based, and some used a few well known ME algorithms.
• AVC Encoder
• Trellis for quantization coeff selection → Trellis is a more general RDOQ since RDOQ’s
• parameter space is a subset of the trellis parameter space. This was used in the AVC
• encoder since the parameter space for selecting quantization coeffs is smaller than that in
• HEVC and hence a higher complexity search algorithm could be afforded. The trade-offs
• that were derived were comparable to that provided by x264, the most popular H264
• encoder.
• VVC Encoder
• In-loop reshaping with local illumination compensation in image coding → Refer USPTO
• Patent Number 11122270 and JVET Contribution Document O0190.
• Realization of RPR based real-time VVC decode and playback on ARM based mobile devices
• → Refer JVET Contribution Document R0224.
• Optimised Encoder Feature Selection
• A greedy algorithm was invented to select a ’good’ point in the encoder parameter space
• that exceeded some specified quality and speed constraints. This has been used for selecting
• optimal parameters in a large sampling of the projects listed above.

Oct 2013 – Apr 2017 · 3 yrs 6 mos

• HEVC Encoder
• CU tree selection algorithm → HEVC provides multiple configurations of CU (Coding
• Unit) trees for signalling in the bitstream. An agglomerative clustering algorithm was used
• to choose sets of CU trees, amongst all available tree configurations, for evaluation and
• selection via a RDO cost function.
• TU tree selection algorithm → HEVC provides multiple configurations of TU (Transform
• Unit) trees every CU for signalling in the bitstream. A heuristics-based algorithm was
• invented and used for selecting a ’good’ TU tree.
• Fast DCI-P3 to Rec-2020 colour-space converter → Because of the enormous disparity
• between the sizes of the co-ordinate systems of the input/output and the transfer function
• domains, and the non-linearity and consequent complexity of the mathematical formulae
• and their software implementation, a novel solution was invented using variable sized
• LUT’s to speed up this conversion.

Jan 2012 – Oct 2013 · 1 yr 9 mos

Aug 2011 – Dec 2011 · 4 mos · Bangalore