|
import os.path as osp |
|
import numpy as np |
|
import mmcv, os |
|
|
|
|
|
|
|
|
|
''' |
|
-------------------------------------------------------------------------------- |
|
This script modifies the original Rellis 3D dataset from 20 labels to 5 labels |
|
|
|
# Original classes |
|
# Labels ID = ["void", "dirt", "grass", "tree", "pole", "water", "sky", "vehicle", |
|
# "object", "asphalt", "building", "log", "person", "fence", "bush", |
|
# "concrete", "barrier", "puddle", "mud", "rubble"] |
|
|
|
# Grouped classes |
|
# New IDs |
|
# 0 -- background: void, sky |
|
# 1 -- obstacle: tree, bush, person, rubble, barrier, log, fence, vehicle, object, pole, building |
|
# 2 -- withwater: mud, puddle, water |
|
# 3 -- unstable: grass, dirt |
|
# 4 -- stable: concrete, asphalt |
|
|
|
-------------------------------------------------------------------------------- |
|
|
|
Python environment requirements |
|
install: pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/cu101/torch1.6.0/index.html |
|
|
|
|
|
|
|
|
|
Download original Rellis 3D: |
|
- RGB images: Rellis_3D_pylon_camera_node.zip (11GB) |
|
Link: https://drive.google.com/file/d/1F3Leu0H_m6aPVpZITragfreO_SGtL2yV/view |
|
- ID Masks : Rellis_3D_pylon_camera_node_label_id.zip (117MB) |
|
Link: https://drive.google.com/file/d/16URBUQn_VOGvUqfms-0I8HHKMtjPHsu5/view |
|
|
|
1. Create the following folders in data/rellis folder |
|
- image |
|
- masks_id |
|
2. Copy Rellis_3D_pylon_camera_node.zip in data/rellis/images folder |
|
3. Copy Rellis_3D_pylon_camera_node_label_id.zip in data/rellis/masks_id folder |
|
4. Rellis_3D_pylon_camera_node_label_color.zip (not used in GanAV model) |
|
5. Unzip Rellis_3D_pylon_camera_node.zip in the data/rellis/images folder |
|
- data/rellis/images/Rellis-3D\00000 |
|
- data/rellis/images/Rellis-3D\00001 |
|
... |
|
6. Move the content of data/rellis/images/Rellis-3D folder to data/rellis/images folder |
|
7. Delete Rellis-3D folder (data/rellis/images/Rellis-3D) |
|
8. Unzip Rellis_3D_pylon_camera_node_label_id.zip in the data/rellis/masks_id folder |
|
- data/rellis/masks_id/Rellis-3D\00000 |
|
- data/rellis/masks_id/Rellis-3D\00001 |
|
... |
|
9. Move the content of data/rellis/masks_id/Rellis-3D folder to data/rellis/masks_id folder |
|
10. Delete Rellis-3D folder (data/rellis/masks_id/Rellis-3D) |
|
11. Detele zip files |
|
12. Copy the files inside the each pylon_camera_node_label_id folder of masks_id folder to masks_id/0000x folder |
|
13. Delete all empty pylon_camera_node_label_id folders |
|
|
|
|
|
''' |
|
rellis_dir=os.getcwd() |
|
|
|
annotation_folder = "/Rellis-3D/masks_id/" |
|
|
|
IDs = [0, 1, 3, 4, 5, 6, 7, 8, 9, 10, 12, 15, 17, 18, 19, 23, 27, 31, 33, 34] |
|
Groups = (0, 3, 3, 1, 1, 2, 0, 1, 1, 4, 1, 1, 1, 1, 1, 4, 1, 2, 2, 1) |
|
|
|
ID_seq = {} |
|
ID_group = {} |
|
for n, label in enumerate(IDs): |
|
ID_seq[label] = n |
|
ID_group[label] = Groups[n] |
|
|
|
|
|
def raw_to_seq(seg): |
|
h, w = seg.shape |
|
out1 = np.zeros((h, w)) |
|
out2 = np.zeros((h, w)) |
|
for i in IDs: |
|
out1[seg==i] = ID_seq[i] |
|
out2[seg==i] = ID_group[i] |
|
|
|
return out1, out2 |
|
|
|
|
|
with open(osp.join(rellis_dir, 'train.txt'), 'r') as r: |
|
i = 0 |
|
for l in r: |
|
print("train: {}".format(i)) |
|
|
|
|
|
file_client_args=dict(backend='disk') |
|
file_client = mmcv.FileClient(**file_client_args) |
|
img_bytes = file_client.get(rellis_dir + annotation_folder + l.strip() + '.png') |
|
gt_semantic_seg = mmcv.imfrombytes(img_bytes, flag='unchanged', backend='pillow').squeeze().astype(np.uint8) |
|
out1, out2 = raw_to_seq(gt_semantic_seg) |
|
|
|
|
|
mmcv.imwrite(out2, rellis_dir + annotation_folder + l.strip() + "_5.png") |
|
|
|
i += 1 |
|
|
|
|
|
with open(osp.join(rellis_dir, 'val.txt'), 'r') as r: |
|
i = 0 |
|
for l in r: |
|
print("val: {}".format(i)) |
|
|
|
|
|
file_client_args=dict(backend='disk') |
|
file_client = mmcv.FileClient(**file_client_args) |
|
img_bytes = file_client.get(rellis_dir + annotation_folder + l.strip() + '.png') |
|
gt_semantic_seg = mmcv.imfrombytes(img_bytes, flag='unchanged', backend='pillow').squeeze().astype(np.uint8) |
|
out1, out2 = raw_to_seq(gt_semantic_seg) |
|
|
|
|
|
mmcv.imwrite(out2, rellis_dir + annotation_folder + l.strip() + "_5.png") |
|
|
|
i += 1 |
|
|
|
|
|
|
|
with open(osp.join(rellis_dir, 'test.txt'), 'r') as r: |
|
i = 0 |
|
for l in r: |
|
print("test: {}".format(i)) |
|
|
|
|
|
file_client_args=dict(backend='disk') |
|
file_client = mmcv.FileClient(**file_client_args) |
|
img_bytes = file_client.get(rellis_dir + annotation_folder + l.strip() + '.png') |
|
gt_semantic_seg = mmcv.imfrombytes(img_bytes, flag='unchanged', backend='pillow').squeeze().astype(np.uint8) |
|
out1, out2 = raw_to_seq(gt_semantic_seg) |
|
|
|
|
|
mmcv.imwrite(out2, rellis_dir + annotation_folder + l.strip() + "_5.png") |
|
|
|
i += 1 |
|
|
|
|
|
|
|
|
|
print("successful") |