the test results do not seem to match the paper

[Carrotor116]

Hi, I clone the source code and download model weight (NNHDRNet.h5, size 121 MB) and test dateset provided in README.md, then I tested and found result are not consistent with the picture demo/peopleStanding.png.

Such as, for test the scene Test/PAPER/PeopleStanding, I modified the model input size (in line) from (3, 256, 256, 6) to (3, 768, 1280, 6) and input resized images.

The result is follow:

The result provided in this repository is follow :

My test results have severe distortion.

Is this test result normal? or is the weight file the optimal weight?

[Carrotor116]

here is my complete test code, cus_test.py file

import argparse
import glob
import os
import time

import cv2
import numpy as np

from HDR import *


def get_test_data_real(images_path):
    # imgs_np = np.zeros([1, 3, 768, 1024, 6])
    file1 = open(os.path.join(images_path, 'exposure.txt'), 'r')
    Lines = file1.readlines()
    t = [float(i) for i in Lines]

    fs = [_ for _ in sorted(glob.glob(os.path.join(images_path, '*.tif')))]
    assert len(fs) == 3
    Xs = []
    for j, f in enumerate(fs):
        # ldr = (cv2.imread(f, -1) / 65535.0).astype(np.float32)
        ldr = (cv2.imread(f, 1) / 256).astype(np.float32)
        # ldr = cv2.resize(ldr, (1024, 768))
        ldr = cv2.cvtColor(ldr, cv2.COLOR_BGR2RGB)
        hdr = ldr ** 2.2 / (2 ** t[j])
        X = np.concatenate([ldr, hdr], axis=-1)
        Xs.append(X)
        # imgs_np[0, j, :, :, :] = X
    hdr = cv2.imread(os.path.join(images_path, 'HDRImg.hdr'), -1)[..., ::-1]
    return Xs, hdr


class SimulateKalantariDataset(object):
    def __init__(self, root):
        self.root = root
        self.scenes = [_ for _ in os.listdir(root) if os.path.isdir(os.path.join(root, _))]
        assert len(self.scenes) > 0, 'invalid root: {}'.format(root)

    def __len__(self):
        return len(self.scenes)

    @staticmethod
    def zoom_image_np(image, height_max, width_max, interpolation=cv2.INTER_CUBIC):
        _min, _max = np.min(image), np.max(image)
        assert len(image.shape) in (2, 3), 'invalid image with shape: {}'.format(image.shape)
        height_ori, width_ori = image.shape[:2]
        hr = 1. * height_max / height_ori
        wr = 1. * width_max / width_ori
        if np.min([hr, wr]) < 1.:
            r = np.min([hr, wr])
            h = np.ceil(r * height_ori)
            w = np.ceil(r * width_ori)
            image = cv2.resize(image, dsize=(int(w), int(h)), interpolation=interpolation)
            if len(image.shape) == 2:
                image = image[..., np.newaxis]
        image = np.clip(image, _min, _max)
        return image

    def _center_crop(self, x, rate=8 * 32):
        x = self.zoom_image_np(x, 1400, 1400)
        h, w, c = x.shape
        crop_h = h // rate * rate
        crop_w = w // rate * rate
        j = int(round((h - crop_h) / 2.))
        i = int(round((w - crop_w) / 2.))
        x = x[max(0, j):min(h, j + crop_h), max(0, i):min(w, i + crop_w), :]
        if x.shape[:2] != (crop_h, crop_w):
            x = cv2.resize(x, (crop_w, crop_h))
        return x

    def __getitem__(self, item):
        Xs, hdr = get_test_data_real(os.path.join(self.root, self.scenes[item]))

        Xs = [self._center_crop(_) for _ in Xs]  # resize image
        hdr = self._center_crop(hdr)  # resize reference hdr

        sdr = np.stack(Xs, axis=0)
        return np.expand_dims(sdr, 0), hdr, self.scenes[item]


class TimeTic(object):
    def __init__(self) -> None:
        self._pre = {None: time.time()}

    def tic(self, tid=None, unit='s') -> float:
        now = time.time()
        delta = (now - self._pre[tid]) if tid in self._pre.keys() else 0.
        self._pre[tid] = now
        if unit == 's':
            return delta
        elif unit == 'ms':
            return delta * 1e3
        else:
            raise RuntimeError('ERROR: do not support unit: {}'.format(unit))


def main(config, model):
    dataset = SimulateKalantariDataset(config.test_data)
    tonemap = cv2.createTonemapReinhard()
    tic = TimeTic()
    times = []
    for idx, (SDR, hdr, name) in enumerate(dataset):
        tic.tic()
        rs = model.predict(SDR)
        times.append(tic.tic())
        out = rs[0]
        print(' process {}/{}: {}, time: {:.6f}'.format(idx, len(dataset), name, times[-1]))

        cv2.imwrite(os.path.join(config.result_dir, 'nhdr_{}.hdr'.format(name)), rs[0][..., ::-1])
        cv2.imwrite(os.path.join(config.result_dir, 'nhdr_{}_ref.hdr'.format(name)), hdr[..., ::-1])
        out = tonemap.process(out.copy())
        cv2.imwrite(os.path.join(config.result_dir, 'nhdr_{}.png'.format(name)), np.uint8(out * 255)[..., ::-1])

    print('avg time: {:.9f}'.format(np.mean(times)))


if __name__ == "__main__":
    parser = argparse.ArgumentParser()

    # Input Parameters
    parser.add_argument('--test_path', type=str, default="Test/EXTRA/001/")
    parser.add_argument('--gpu', type=int, default=1)
    parser.add_argument('--weight_test_path', type=str, default="weights/best.h5")
    parser.add_argument('--filter', type=int, default=32)
    parser.add_argument('--kernel', type=int, default=3)
    parser.add_argument('--encoder_kernel', type=int, default=3)
    parser.add_argument('--decoder_kernel', type=int, default=4)
    parser.add_argument('--triple_pass_filter', type=int, default=256)
    parser.add_argument('--test_data', type=str, required=True)
    parser.add_argument('--result_dir', type=str, required=True)

    config = parser.parse_args()

    if not os.path.exists(config.result_dir):
        os.mkdir(config.result_dir)

    os.environ['CUDA_VISIBLE_DEVICES'] = str(config.gpu)

    model_x = NHDRRNet(config)
    # x = Input(shape=(3, 256, 256, 6))
    x = Input(shape=(3, 768, 1280, 6))
    out = model_x.main_model(x)
    model = Model(inputs=x, outputs=out)
    model.load_weights(config.weight_test_path)
    model.summary()

    main(config, model)

test command

$ tree Test/PAPER/
Test/PAPER/
├── BarbequeDay
│   ├── 262A2943.tif
│   ├── 262A2944.tif
│   ├── 262A2945.tif
│   ├── exposure.txt
│   └── HDRImg.hdr
├── LadySitting
│   ├── 262A2705.CR2
│   ├── 262A2705.pgm
│   ├── 262A2705.tif
│   ├── 262A2706.CR2
│   ├── 262A2706.pgm
│   ├── 262A2706.tif
│   ├── 262A2707.CR2
│   ├── 262A2707.pgm
│   ├── 262A2707.tif
│   ├── exposure.txt
│   └── HDRImg.hdr
├── ManStanding
│   ├── 262A2629.tif
│   ├── 262A2630.tif
│   ├── 262A2631.tif
│   ├── exposure.txt
│   └── HDRImg.hdr
├── PeopleStanding
│   ├── 262A2866.tif
│   ├── 262A2867.tif
│   ├── 262A2868.tif
│   ├── exposure.txt
│   ├── HDRImg.hdr
└── PeopleTalking
    ├── 262A2810.tif
    ├── 262A2811.tif
    ├── 262A2812.tif
    ├── exposure.txt
    └── HDRImg.hdr

5 directories, 31 files

$ python cus_test.py --weight_test_path NNHDRNet.h5 --test_data Test/PAPER/ --result_dir results --gpu -1

[zhengchaobing]

Hi, I clone the source code and download model weight (NNHDRNet.h5, size 121 MB) and test dateset provided in README.md, then I tested and found result are not consistent with the picture demo/peopleStanding.png.

Such as, for test the scene Test/PAPER/PeopleStanding, I modified the model input size (in line) from (3, 256, 256, 6) to (3, 768, 1280, 6) and input resized images.

The result is follow:

The result provided in this repository is follow :

My test results have severe distortion.

Is this test result normal? or is the weight file the optimal weight?

can you share the weight me?