Function prepare_batch_render

script/feature/direct_feature_matching.py:144–176

Break batch of images into rays

(args, pose, batch_size, target_, H, W, focal, half_res=True, rand=True)

    return pose_loss

def prepare_batch_render(args, pose, batch_size, target_, H, W, focal, half_res=True, rand=True):
    ''' Break batch of images into rays '''
    target_ = target_.permute(0, 2, 3, 1).numpy() # convert to numpy image
    if half_res:
        N_rand = batch_size * (H//2) * (W//2)
        target_half = np.stack([cv2.resize(target_[i], (H//2, W//2), interpolation=cv2.INTER_AREA) for i in range(batch_size)], 0)
        target_half = torch.Tensor(target_half)
        
        rays = torch.stack([torch.stack(get_rays(H//2, W//2, focal/2, pose[i]), 0) for i in range(batch_size)], 0) # [N, ro+rd, H, W, 3] (130, 2, 100, 100, 3)
        rays_rgb = torch.cat((rays, target_half[:, None, ...]), 1)

    else:
        # N_rand = batch_size * H * W
        N_rand = args.N_rand
        target_ = torch.Tensor(target_)
        rays = torch.stack([torch.stack(get_rays(H, W, focal, pose[i]), 0) for i in range(batch_size)], 0) # [N, ro+rd, H, W, 3] (130, 2, 200, 200, 3)
        # [N, ro+rd+rgb, H, W, 3]
        rays_rgb = torch.cat([rays, target_[:, None, ...]], 1)

    # [N, H, W, ro+rd+rgb, 3]
    rays_rgb = rays_rgb.permute(0, 2, 3, 1, 4)
    
    # [(N-1)*H*W, ro+rd+rgb, 3]
    rays_rgb = torch.reshape(rays_rgb, (-1, 3, 3))

    if 1:
        rays_rgb = rays_rgb[torch.randperm(rays_rgb.shape[0])]

    # Random over all images
    batch = rays_rgb[:N_rand].permute(1, 0 , 2) # [B, 2+1, 3*?] # (4096, 3, 3)
    batch_rays, target_s = batch[:2], batch[2] # [2, 4096, 3], [4096, 3]

    return batch_rays, target_s

def eval_on_batch(args, data, model, feat_model, pose, img_idx, hwf, half_res, device, world_setup_dict, **render_kwargs_test):
    ''' Perform 1 step of eval'''