This paper introduces a novel convolutional network architecture, the "stacked hourglass," designed for human pose estimation. The architecture processes features across all scales and consolidates them to capture various spatial relationships in the body. The key innovation is the repeated bottom-up and top-down processing, combined with intermediate supervision, which significantly improves the network's performance. The stacked hourglass network is evaluated on the FLIC and MPII benchmarks, achieving state-of-the-art results with over 2% improvement in average accuracy across all joints and up to 4-5% improvement on challenging joints like the knees and ankles. The paper also discusses related work, network architecture details, training methods, and ablation experiments, highlighting the effectiveness of the stacked hourglass design and intermediate supervision in handling complex pose estimation tasks.This paper introduces a novel convolutional network architecture, the "stacked hourglass," designed for human pose estimation. The architecture processes features across all scales and consolidates them to capture various spatial relationships in the body. The key innovation is the repeated bottom-up and top-down processing, combined with intermediate supervision, which significantly improves the network's performance. The stacked hourglass network is evaluated on the FLIC and MPII benchmarks, achieving state-of-the-art results with over 2% improvement in average accuracy across all joints and up to 4-5% improvement on challenging joints like the knees and ankles. The paper also discusses related work, network architecture details, training methods, and ablation experiments, highlighting the effectiveness of the stacked hourglass design and intermediate supervision in handling complex pose estimation tasks.