Human Mesh Estimation
SMPL[4] is a statistical model which allows forming a human mesh using 10 parameters to define the body shape and 72 parameters to define the body pose. Human mesh estimation methods such as HMR[5], SPIN[6] estimate SMPL[4] parameters by passing the image to a CNN model. SPIN[6] also uses an iterative optimization loop to fit the SMPL mesh to 2D key points estimated on the RGB image. CRA[7] estimates the human mesh parameters in a feed forward way by using multiple intermediate representations, specifically, using 2D keypoints and UV maps. 2D key points provide sparse but robust information, while UV maps provide dense but noisy information. These intermediate representations are used to provide complimentary information for the model to estimate the human mesh.
Pressure Prediction
In Pressure Vision[8], the authors estimate pressure from a single RGB image, using an encoder-decoder stack, giving a per-pixel pressure output. In Visual Haptic Reasoning[9], the authors estimate pressure on points on the human body. They train their models on simulated datasets and show that the pressure estimation helps in tasks such as covering the human arm with cloth.
Contact Pressure Prediction
In contrast to human mesh estimation for generalized settings, such as pedestrians crossing roads, fewer works have focussed on resting poses. These scenarios cover poses where humans are lying down or sitting and have contact b/w the surface and body. The task for rest pose estimation, especially for medical scenarios, is challenging due to limitation of not using RGB images and occlusion from blankets.
Figure 4: Heavy Occlusion examples: Depth Image Modality
BodyPressure[10] uses a residual learning method to regress the SMPL parameters from a depth image. This approach builds the baseline for our work, where first human pose is estimated and then pressure is mapped on the estimated body.