🩹 [Fix] Labels in YOLO detection format

[luuzk]

This PR fixes a bug where .txt files are always interpreted in YOLO segmentation format (class_id, x1, y1, x2, y2, ..., xn, yn) although they are in very common YOLO detection format (class_id, cx, cy, w, h). This behavior is explained in #141 (comment) and #148 (comment) by @henrytsui000.

Closes #102, #141, #148, and #158.

Adds a new method convert_bboxes to convert bounding boxes to YOLO segmentation format if they are in YOLO detection format. Otherwise, leave them as is for backwards compatibility.

[minnemint]

You saved my life, thank you

[PyBastian]

Would this mean that the bbox from yolo format, its consider as a 4 points segmentation for the model?

[divyanidhi]

Hi @luuzk
Adding this PR to code does solve the issue of yolo format, but I am not able to get correct inferencing on the custom trained model. Adding image as a reference.

Do you have any suggestions on what could be the issue?

[johnk2hawaii]

Thanks for this, I have tried this branch and I am still getting the same problem as always, when I transfer train it ends in a couple of seconds with no error report and a one liner in my output log that says

📈 Enable Model EMA

I have task.ema.enable=True but if I set it to False nothing changes. Working in google colab if that helps.

Cheers

[johnk2hawaii]

I made some progress by altering the dataset folder structure but the error I am getting now is:

Error executing job with overrides: ['task=train', 'task.data.batch_size=5', 'task.ema.enable=True', 'model=v9-s', 'dataset=data.yaml', 'device=cuda', 'task.epoch=10', 'use_wandb=False'] Traceback (most recent call last): File "C:\Users\johnk\Desktop\pythonProject\pythonProject2\MITCorrection\YOLO\yolo\lazy.py", line 34, in main model = TrainModel(cfg) File "C:\Users\johnk\Desktop\pythonProject\pythonProject2\MITCorrection\YOLO\yolo\tools\solver.py", line 70, in __init__ super().__init__(cfg) File "C:\Users\johnk\Desktop\pythonProject\pythonProject2\MITCorrection\YOLO\yolo\tools\solver.py", line 35, in __init__ self.val_loader = create_dataloader(self.validation_cfg.data, self.cfg.dataset, self.validation_cfg.task) File "C:\Users\johnk\Desktop\pythonProject\pythonProject2\MITCorrection\YOLO\yolo\tools\data_loader.py", line 225, in create_dataloader dataset = YoloDataset(data_cfg, dataset_cfg, task) File "C:\Users\johnk\Desktop\pythonProject\pythonProject2\MITCorrection\YOLO\yolo\tools\data_loader.py", line 39, in __init__ self.img_paths, self.bboxes, self.ratios = tensorlize(self.load_data(Path(dataset_cfg.path), phase_name)) File "C:\Users\johnk\Desktop\pythonProject\pythonProject2\MITCorrection\YOLO\yolo\utils\dataset_utils.py", line 127, in tensorlize raise e File "C:\Users\johnk\Desktop\pythonProject\pythonProject2\MITCorrection\YOLO\yolo\utils\dataset_utils.py", line 121, in tensorlize img_paths, bboxes, img_ratios = zip(*data) ValueError: not enough values to unpack (expected 3, got 0)
It looks like the data object is empty. My dataset is in YOLO format looks like this

Xray/
├── images
│ ├── train
│ └── val
├── labels
│ ├── train
│ └── val

my data.yaml file is:

path: /content/YOLO/yolo/config/dataset/data.yaml
train: /content/YOLO/data/Xray/images/train
validation: /content/YOLO/data/Xray/images/valid

train_label_dir: /content/YOLO/data/Xray/labels/train
val_label_dir: /content/YOLO/data/Xray/labels/valid

class_num: 5
names: ['0', '1', '2', '3', '4']

Hoping to get some advice.

[pypoulp]

Hello,

Currently on main, the code accepts annotations in the format (class_id, x1, y1, x2, y2) as well.

Attempting automatic conversion based on:

if len(anno) > 5:
    segmentation_data.append(anno)
    continue

breaks support for this format.

I would suggest introducing a configuration option, allowing users to explicitly specify the annotation format. This would make the behavior clearer and avoid unexpected issues.

[jenhaoyang]

Use YOLOv8 Format
https://docs.ultralytics.com/datasets/segment/