why the mean_t and std_t for 7-Scenes are set to zeros(3) and ones(3) in SevenScenes dataset?

[ZhouJiaHuan]

Hey, I noticed that the pose stats of 7-Scenes, namely the mean_t and std_t are simply set to zeros(3) and ones(3). But when I use the pose stats computed by myself (not equal to zeros(3) and ones(3)) for training and evaluation, the accuracy decreased a lot.

[samarth-robo]

Hi @ZhouJiaHuan. IIRC the mean_t and std_t were only used to avoid making the network predict super large values. This applies to RobotCar which is a large environment, but not so much to 7Scenes. Hence they are set to identity.
About your problem: did you verify that you can train and test properly if you use the identity pose stats?

[ZhouJiaHuan]

Thanks for your replying. @samarth-robo

Yea, because my environment is Python3 + PyTorch 1.5.1, I modified some codes for compatibility and then trained first with all the default parameters (including the identity pose stats). The val result is almost same with the paper. The experimental result is as follows:

---

• model: MapNet
• Dataset: 7-Scenes Heads
• Error in translation: median 0.19 m (0.18 m in paper TABLE 3), mean 0.23 m
• Error in rotation: median 13.23 degrees (13.25 degrees in paper TABLE 3), mean 13.26 degree

---

I am not sure whether the different PyTorch version will affect the result a lot. (the pretrained weights of ResNet-34 in Torch 1.5.1 is different with that in Torch 0.4.1)

However, when I substitute the identity pose stats with the value computed from Heads training data, the accuracy dropped a lot. The experimental result is as follows:

---

• model: MapNet
• Dataset: 7-Scenes Heads
• mean_t: (-0.01829027, -0.13756783, 0.09475311)
• std_t: (0.42199861, 0.19139078, 0.13966574)
• Error in translation: median 0.95 m (0.18 m in paper TABLE 3), mean 0.99 m
• Error in rotation: median 14.47 degrees (13.25 degrees in paper TABLE 3), mean 13.91 degree

---

And as a reference, I also visualize the training process with TensorboardX

The figure above is the learning parameters in MapNet loss function and the figure below is the training/val loss. As you see, the learning parameters are almost same during training pipeline. But the training loss for computed pose stats (in blue) seems not converge well.

[samarth-robo]

@ZhouJiaHuan OK. And just to confirm, you calculated pose stats similarly to RobotCar, by modifying L100 and L101 of seven_scenes.py as follows?

poses = np.vstack(ps.values())
mean_t = np.mean(poses[:, [3, 7, 11]], axis=0)
std_t = np.std(poses[:, [3, 7, 11]], axis=0)

[ZhouJiaHuan]

@samarth-robo That's my code to calculate pose stats.

import os
import os.path as osp
import numpy as np
import argparse

DATA_DIR = "data/deepslam_data/"


def parse_args():
    parser = argparse.ArgumentParser(description='Dataset pose statistics')
    parser.add_argument('--dataset', type=str, choices=('SevenScenes', 'UnoRobot'),
                        help='Dataset', required=True)
    parser.add_argument('--scene', type=str,
                        help='Scene name', required=True)
    args = parser.parse_args()
    return args


def parse_poses(dataset, scene):
    data_path = osp.join(DATA_DIR, dataset, scene)
    assert osp.exists(data_path)
    split_file = osp.join(data_path, 'TrainSplit.txt')
    assert osp.exists(split_file)
    with open(split_file, 'r') as f:
        seqs = [int(l.split('sequence')[-1]) for l in f if not l.startswith('#')]
    pss = []
    for seq in seqs:
        seq_dir = osp.join(data_path, 'seq-{:02d}'.format(seq))
        p_filenames = [n for n in os.listdir(osp.join(seq_dir, '.')) if
                       n.find('pose') >= 0]

        frame_idx = np.array(range(len(p_filenames)), dtype=np.int)
        ps = [np.loadtxt(osp.join(seq_dir, 'frame-{:06d}.pose.txt'.
              format(i)))[:3, 3] for i in frame_idx]
        pss.extend(ps)
    return np.array(pss)


if __name__ == "__main__":
    args = parse_args()
    pss_array = parse_poses(args.dataset, args.scene)
    print(pss_array.shape)
    pss_mean = np.mean(pss_array, axis=0)
    pss_std = np.std(pss_array, axis=0)
    print("pose mean = {}".format(pss_mean))
    print("pose std = {}".format(pss_std))`

the code below is equivalent to that in RobotCar:

ps = [np.loadtxt(osp.join(seq_dir, 'frame-{:06d}.pose.txt'. format(i)))[:3, 3] for i in frame_idx]
pss.extend(ps)

[ZhouJiaHuan]

@samarth-robo hey, I found the causes. It's my inference codes problem. I forgot to change the pose stats during inference phase so that the pose stats is still the identity. I fixed this error and the error for mapnet on heads is 0.20 m, 13.91 degree. Pardon the bother and thanks for your kindness.

[samarth-robo]

@ZhouJiaHuan I was just about to comment that! Glad you found the solution. I'll close this issue now.