Acceleration Error.

[Songinpyo]

First of all, thanks for your amazing work.

I have a question about acceleration error of METRO human mesh recovery model.

I calculated that on 3DPW test set.

Although this is actually for image based model ( did not focusing on smoothness over the frames ) the value is too high.
I think I did some mistakes on calculating.

So I would like to ask you about

1. Did you calculated the accel error of METRO? Is there any codes officialy supporting?
2. Are my codes snnipets for calcuating accel error is wrong?

def compute_error_accel(joints_gt, joints_pred, vis=None):

    joints_gt = joints_gt[vis == 1]
    # joints_gt = joints_gt[:, :, :-1]
    joints_pred = joints_pred[vis == 1]
    
    gt_pelvis = (joints_gt[:, 2,:] + joints_gt[:, 3,:]) / 2
    joints_gt = joints_gt - gt_pelvis[:, None, :]
    pred_pelvis = (joints_pred[:, 2,:] + joints_pred[:, 3,:]) / 2
    joints_pred = joints_pred - pred_pelvis[:, None, :]
    # (N-2)x14x3
    accel_gt = joints_gt[:-2] - 2 * joints_gt[1:-1] + joints_gt[2:]
    accel_pred = joints_pred[:-2] - 2 * joints_pred[1:-1] + joints_pred[2:]

    normed = np.linalg.norm(accel_pred - accel_gt, axis=2)
   
    return np.mean(normed, axis=1)

And from validate code

    pred_j3ds_all, gt_j3ds_all, vis_j3ds_all = [], [], []
    
    # switch to evaluate mode
    METRO_model.eval()
    smpl.eval()
    with torch.no_grad():
        for i, (img_keys, images, annotations) in enumerate(val_loader):                   
            batch_size = images.size(0)
            # compute output
            images = images.cuda(args.device)
            gt_3d_joints = annotations['joints_3d'].cuda(args.device)
            gt_3d_pelvis = gt_3d_joints[:,cfg.J24_NAME.index('Pelvis'),:3]
            gt_3d_joints = gt_3d_joints[:,cfg.J24_TO_J14,:]
            gt_3d_joints[:,:,:3] = gt_3d_joints[:,:,:3] - gt_3d_pelvis[:, None, :]
            has_3d_joints = annotations['has_3d_joints'].cuda(args.device)

            gt_pose = annotations['pose'].cuda(args.device)
            gt_betas = annotations['betas'].cuda(args.device)
            has_smpl = annotations['has_smpl'].cuda(args.device)

            # generate simplified mesh
            gt_vertices = smpl(gt_pose, gt_betas)
            gt_vertices_sub = mesh_sampler.downsample(gt_vertices)
            gt_vertices_sub2 = mesh_sampler.downsample(gt_vertices_sub, n1=1, n2=2)

            # normalize gt based on smpl pelvis 
            gt_smpl_3d_joints = smpl.get_h36m_joints(gt_vertices)
            gt_smpl_3d_pelvis = gt_smpl_3d_joints[:,cfg.H36M_J17_NAME.index('Pelvis'),:]
            gt_vertices_sub2 = gt_vertices_sub2 - gt_smpl_3d_pelvis[:, None, :] 
            gt_vertices = gt_vertices - gt_smpl_3d_pelvis[:, None, :] 

            # forward-pass
            pred_camera, pred_3d_joints, pred_vertices_sub2, pred_vertices_sub, pred_vertices = METRO_model(images, smpl, mesh_sampler)
            
            # obtain 3d joints from full mesh
            pred_3d_joints_from_smpl = smpl.get_h36m_joints(pred_vertices)

            pred_3d_pelvis = pred_3d_joints_from_smpl[:,cfg.H36M_J17_NAME.index('Pelvis'),:]
            pred_3d_joints_from_smpl = pred_3d_joints_from_smpl[:,cfg.H36M_J17_TO_J14,:]
            pred_3d_joints_from_smpl = pred_3d_joints_from_smpl - pred_3d_pelvis[:, None, :]
            pred_vertices = pred_vertices - pred_3d_pelvis[:, None, :]

            pred_j3ds_all.append(pred_3d_joints_from_smpl.cpu().numpy())
            gt_j3ds_all.append(gt_3d_joints[:,:,:3].cpu().numpy())
            vis_j3ds_all.append(has_3d_joints.cpu().numpy())

    pred_j3ds_all = np.concatenate(pred_j3ds_all)
    gt_j3ds_all = np.concatenate(gt_j3ds_all)
    vis_j3ds_all = np.concatenate(vis_j3ds_all)
    accel_error = compute_error_accel(joints_gt=gt_j3ds_all, joints_pred=pred_j3ds_all, vis=vis_j3ds_all) # m to mm
    mean_accel_error = np.mean(np.concatenate(accel_error))
    print("Mean Accel Error :", mean_accel_error * 1000)

I'll wait for your answering.
Thanks!

[ZeyuPing]

Hi! could u please offer me the pretrained model file for me?