load_EV2.py

import os
import torch
import numpy as np
import imageio 
import json
import torch.nn.functional as F
import cv2

from utils import get_bbox3d_for_EV

trans_t = lambda t : torch.Tensor([
    [1,0,0,0],
    [0,1,0,0],
    [0,0,1,t],
    [0,0,0,1]]).float()

rot_phi = lambda phi : torch.Tensor([
    [1,0,0,0],
    [0,np.cos(phi),-np.sin(phi),0],
    [0,np.sin(phi), np.cos(phi),0],
    [0,0,0,1]]).float()

rot_theta = lambda th : torch.Tensor([
    [np.cos(th),0,-np.sin(th),0],
    [0,1,0,0],
    [np.sin(th),0, np.cos(th),0],
    [0,0,0,1]]).float()


def pose_spherical(theta, phi, radius):
    c2w = trans_t(radius)
    c2w = rot_phi(phi/180.*np.pi) @ c2w
    c2w = rot_theta(theta/180.*np.pi) @ c2w
    c2w = torch.Tensor(np.array([[-1,0,0,0],[0,0,1,0],[0,1,0,0],[0,0,0,1]])) @ c2w
    return c2w


def load_EV_data(basedir, half_res=False, testskip=1):
    splits = ['train', 'val', 'test']
    metas = {}
    for s in splits:
        with open(os.path.join(basedir, 'transforms_{}.json'.format(s)), 'r') as fp:
            metas[s] = json.load(fp)

    all_imgs = []
    all_poses = []
    all_intrinsics = []
    counts = [0]
    for s in splits:
        meta = metas[s]
        imgs = []
        poses = []
        intrinsics = []
        if s=='train' or testskip==0:
            skip = 1
        else:
            skip = testskip
            
        for frame in meta['frames'][::skip]:
            fname = os.path.join(basedir, frame['file_path'][2:])
            imgs.append(imageio.imread(fname))
            poses.append(np.array(frame['transform_matrix']))
            
            focal = frame['fl_x']
            cx = frame['cx']
            cy = frame['cy']
            
            K = np.array([[focal, 0, cx], [0, focal, cy], [0, 0, 1] ])
            intrinsics.append(K)
        
        
        for i in range(len(imgs)): imgs[i] = (imgs[i] / 255.).astype(np.float32)
        #imgs = (np.array(imgs) / 255.).astype(np.float32) # keep all 4 channels (RGBA)
        
        poses = np.array(poses).astype(np.float32)
        intrinsics = np.array(intrinsics).astype(np.float32)
        
        counts.append(counts[-1] + len(imgs))
        # counts.append(counts[-1] + imgs.shape[0])
        
        all_imgs.append(imgs)
        all_poses.append(poses)
        all_intrinsics.append(intrinsics)
    
    i_split = [np.arange(counts[i], counts[i+1]) for i in range(3)]
    
    imgs = [item for sublist in all_imgs for item in sublist]
    #imgs = np.concatenate(all_imgs, 0)
    poses = np.concatenate(all_poses, 0)
    intrinsics = np.concatenate(all_intrinsics, 0)
    
    # H, W = imgs[0].shape[:2]
    # camera_angle_x = float(meta['camera_angle_x'])
    # focal = .5 * W / np.tan(.5 * camera_angle_x)
    
    render_poses = torch.stack([pose_spherical(angle, -30.0, 4.0) for angle in np.linspace(-180,180,10+1)[:-1]], 0)
    
    # if half_res:
    #     H = H//2
    #     W = W//2
    #     focal = focal/2.

    #     imgs_half_res = np.zeros((imgs.shape[0], H, W, 4))
    #     for i, img in enumerate(imgs):
    #         imgs_half_res[i] = cv2.resize(img, (W, H), interpolation=cv2.INTER_AREA)
    #     imgs = imgs_half_res
    #     # imgs = tf.image.resize_area(imgs, [400, 400]).numpy()

    bounding_box = get_bbox3d_for_EV(metas["train"], near=2.0, far=6.0)

    return imgs, poses, render_poses, intrinsics, i_split, bounding_box