PCHMR
A benchmark for human mesh recovery (HMR) under privacy constraints. It evaluates, based on federated learning and secure aggregation, the performance degradation of HMR models in scenarios where data is not centralized.
Timeline (last 90 days)
PCHMR is a benchmark for human mesh recovery (HMR) under privacy constraints, aimed at evaluating and improving HMR models without centralizing raw images.
PCHMR uses federated learning and a secure aggregation variant to train state-of-the-art HMR models, and evaluates their performance across a range of client scales and degrees of data heterogeneity.
Neither the federated learning nor the secure aggregation training path of PCHMR centralizes raw images.
PCHMR deliberately varies client scale and degree of data heterogeneity to measure the impact of distribution shift on the HMR backbone network.
All Facts (4)
PCHMR is a benchmark for human mesh recovery (HMR) under privacy constraints, aimed at evaluating and improving HMR models without centralizing raw images.
50%UnverifiedPCHMR uses federated learning and a secure aggregation variant to train state-of-the-art HMR models, and evaluates their performance across a range of client scales and degrees of data heterogeneity.
50%UnverifiedNeither the federated learning nor the secure aggregation training path of PCHMR centralizes raw images.
50%UnverifiedPCHMR deliberately varies client scale and degree of data heterogeneity to measure the impact of distribution shift on the HMR backbone network.
50%