Converts txt kitti labels to yolo v2 labels.
Path of the folders where the images and labels are saved, and the output directory
pickle files containing a dictionary with the image and its corresponding labels for yolo v2.
file#.p: {'image':image,'label':labels}The output labels are contained on a numpy array with zero shape of [13,13,5,6], where the only non zeros are those anchors with an object in them.
NOTE: It will create a file for each image.
First of all, we have to make sure that the number of images and labels files are the same, and the order of the images correspond to the order of the labels. For example:
OR OR OR
Folder_1 | Folder_1 | Folder_1 | Folder_1
. | . | . | .
├─ im1.jpg | ├─ im1.jpg | ├─ im1.jpg | ├─ im1.jpg
├─ im2.jpg | ├─ im2.jpg | ├─ im2.jpg | ├─ im2.jpg
├─ im3.jpg | ├─ im3.jpg | ├─ im3.jpg | ├─ im3.jpg
└─ im4.jpg | └─ im4.jpg | ├─ im4.jpg | ├─ im4.jpg
| | ├─ lb1.txt | ├─ asd.tx
Folder_2 | Folder_2 | ├─ lb2.txt | ├─ bda.txt
. | . | ├─ lb3.txt | ├─ crew.txt
├─ lb1.txt | ├─ asd.txt | └─ lb4.txt | └─ das.txt
├─ lb2.txt | ├─ bda.txt |
├─ lb3.txt | ├─ crew.txt |
└─ lb4.txt | └─ das.txt |
You get the idea.
from data_transform import Kitti2Yolo
ims_folder = '../ims/'
lbs_folder = '../lbs/'
out_folder = '../data/'
conv = Kitti2Yolo(ims_folder,lbs_folder)
conv.convert(out_dir=out_folder)- Add saving support for multiple class. Right now it saves all the images with the class label 15, which indicates a person. It's not hard, it actually already reads the class on the file, it just need to compare it to some list containing the right class and assign it.
- Dump multiple images/labels into a single file.
Adaptation from https://github.com/pmkalshetti/object_detection/blob/master/src/write_data_to_TFRecords.py . (I wasn't able to use tfrecords correctly :P)