2 offres trouvées

Summary In this internship at the London AI Video Lab, the objective is to design computationally efficient video decoders in an AI-based video compression codec. Current AI-based video compression models outperform conventional codecs, like HEVC, VVC and AV1. However, this comes at the cost of impractical compute requirements: at decode, current AI-based video compression decoders are several orders of magnitude more complex than conventional video compression decoders. The goal of the internship is to design efficient AI-based decoders that leverage spatial sparsity to reduce their computational complexity. This work will be seen as one step forward toward the deployment of end-to-end trained AI-based video compression models. The goal will be to find and review potential existing methods of spatial sparsity in AI-based video models. In a second step, spatially sparse AI-based decoders will be designed, implemented and integrated into the London AI Video Lab's end-to-end...

Summary In this internship at the London AI Video Lab, the objective is to study fixed-point arithmetic solutions for ensuring bit-exact video compression in AI-based video codecs. Current AI-based video compression models outperform conventional codecs, like HEVC, VVC and AV1. However, AI-based video compression models are trained using floating-point arithmetic. Unfortunately, floating point arithmetic is insufficient to ensure bit-exact execution. Bit-exact execution is needed to ensure encoded bitstreams are universally decodable across any device. Fixed-point arithmetic is a potential solution to this problem. The goal of the internship is to determine a fixed-point arithmetic setup capable of ensuring bit-exactness while maintaining model performance. This work will be seen as one step forward toward the deployment of end-to-end trained AI-based video compression models. The goal will be to study various fixed-point arithmetic setups for layers and components of AI-based...