Do I need to disable image flipping/mirroring?

adrianholovaty commented 3 years ago

Hello — thanks so much for this project. It's really great! Not only is the performance top-notch, the documentation and developer experience are very solid as well.

I noticed in the generated "mosaic" images that sometimes images are flipped into a mirror image. Does this mean that yolo is learning to detect objects in that flipped perspective? I work with image data in which there is an important difference between an image and the "flipped" version of itself, so I want to make sure my model won't be learning the wrong thing.

If yolo is indeed learning the flipped images, is there a way to disable this? Or do I not need to worry about it?

github-actions[bot] commented 3 years ago

👋 Hello @adrianholovaty, thank you for your interest in 🚀 YOLOv5! Please visit our ⭐️ Tutorials to get started, where you can find quickstart guides for simple tasks like Custom Data Training all the way to advanced concepts like Hyperparameter Evolution.

If this is a 🐛 Bug Report, please provide screenshots and minimum viable code to reproduce your issue, otherwise we can not help you.

If this is a custom training ❓ Question, please provide as much information as possible, including dataset images, training logs, screenshots, and a public link to online W&B logging if available.

For business inquiries or professional support requests please visit https://www.ultralytics.com or email Glenn Jocher at glenn.jocher@ultralytics.com.

Requirements

Python 3.8 or later with all requirements.txt dependencies installed, including torch>=1.7. To install run:

$ pip install -r requirements.txt

Environments

YOLOv5 may be run in any of the following up-to-date verified environments (with all dependencies including CUDA/CUDNN, Python and PyTorch preinstalled):

Google Colab and Kaggle notebooks with free GPU:
Google Cloud Deep Learning VM. See GCP Quickstart Guide
Amazon Deep Learning AMI. See AWS Quickstart Guide
Docker Image. See Docker Quickstart Guide

Status

If this badge is green, all YOLOv5 GitHub Actions Continuous Integration (CI) tests are currently passing. CI tests verify correct operation of YOLOv5 training (train.py), testing (test.py), inference (detect.py) and export (export.py) on MacOS, Windows, and Ubuntu every 24 hours and on every commit.

mfruhner commented 3 years ago

Hi, yes this means that the model will learn the flipped version as the same class. To disable this you can create your own hyperparamter file (hyp.yaml) and set fliplr: 0.0, so that the probability of the image being flipped while augmenting is zero.

adrianholovaty commented 3 years ago

@mfruhner Thanks very much!

Transigent commented 3 years ago

Hi, yes this means that the model will learn the flipped version as the same class. To disable this you can create your own hyperparamter file (hyp.yaml) and set fliplr: 0.0, so that the probability of the image being flipped while augmenting is zero.

@mfruhner Just a followup if I may, my dataset consists of images and their horizontally mirrored versions that I have added as augmentations.

Does this mean that withfliplr: 0.5 # image flip left-right (probability) 50% of images are also flipped and effectively used as training images during training?
If I set this to 1.0 does that mean for every supplied image its mirror will also be used in training? ie. I dont have to supply the mirrored source images at all?
With something like degrees: 0.0 # image rotation (+/- deg) does that mean that each training image is rotated the specified amount in addition to the unrotated training image? Or instead? Is it a fixed amount or a range for randomization?

Thanks for any guidance.

Transigent commented 3 years ago

PS I read 'Hyperparameter Evolution' but either it didn't have an answer, or I didn't understand well enough to extract one.

glenn-jocher commented 3 years ago

@Transigent the hyperparameter file does not define individual augmentations, it defines distributions which are randomly sampled at runtime. No image is ever seen twice during training. fliplr 0.5 defines a 50% lr flip probability.

Transigent commented 3 years ago

Excellent thanks so much @glenn-jocher .

SyedHamza0196 commented 3 years ago

@glenn-jocher what about " degrees: 0.0 # image rotation (+/- deg) " . Is it probability or 0.5 means 90 degrees rotation

glenn-jocher commented 3 years ago

@SyedHamza0196 rotation units are in comment # image rotation (+/- deg)

chinhcd commented 2 years ago

@adrianholovaty: Hello, did you try to disable the flip function in the hyperparameter? does it work? I set the fliplr = 0 (in data/hyps/hyp.scratch.yaml) but the model after training still detects the flip left-right image.

chinhcd commented 2 years ago

@glenn-jocher : I set the fliplr = 0, but after training, the model still detects the flip left-right image. Is there another parameter I need to change to disable the flip images function? ( flipud is disabled) Thank you very much.

glenn-jocher commented 2 years ago

@chinhcd fliplr hyperparameter control this augmentation. There is no other parameter.

chinhcd commented 2 years ago

@glenn-jocher : the fliplr setting in hyperparameter is strange. when you have time, could you please check it. I trained a model with 2 classes: flip and non-flip (lr). the mAP0.5 is good(0.996) but when I tested with both train and val images, the model detected wrong (about 50%) it happens with fliplr only.

In addition, with the model (fliplr = 0.0), if we test with the flipup images and fliplr images, only fliplr images always return high confidence score.

Thank you very much.

glenn-jocher commented 2 years ago

@chinhcd I have extremely little time, but if you create a reproducible example of a possible problem our team can start to investigate. With the above information we have nothing to reproduce.

sarpx commented 2 years ago

@glenn-jocher I've had this problem myself. Even though I did fliplr 0.0, the results worked in the opposite direction of what I wanted. Now I'm making fliplr 1.0 and giving another training, but I'm not so sure about the result. It looks like we can't turn off right and left rotation. This is very important for some trainings. The opposite of something can look like another letter or another object that we teach, and it causes confusion.

glenn-jocher commented 2 years ago

@sarpx 👋 Hello! Thanks for asking about image augmentation. YOLOv5 🚀 applies online imagespace and colorspace augmentations in the trainloader (but not the val_loader) to present a new and unique augmented Mosaic (original image + 3 random images) each time an image is loaded for training. Images are never presented twice in the same way.

YOLOv5 augmentation

Augmentation Hyperparameters

The hyperparameters used to define these augmentations are in your hyperparameter file (default data/hyp.scratch.yaml) defined when training:

python train.py --hyp hyp.scratch-low.yaml

https://github.com/ultralytics/yolov5/blob/b94b59e199047aa8bf2cdd4401ae9f5f42b929e6/data/hyps/hyp.scratch-low.yaml#L6-L34

Augmentation Previews

You can view the effect of your augmentation policy in your train_batch*.jpg images once training starts. These images will be in your train logging directory, typically yolov5/runs/train/exp:

train_batch0.jpg shows train batch 0 mosaics and labels:

YOLOv5 Albumentations Integration

YOLOv5 🚀 is now fully integrated with Albumentations, a popular open-source image augmentation package. Now you can train the world's best Vision AI models even better with custom Albumentations 😃!

PR https://github.com/ultralytics/yolov5/pull/3882 implements this integration, which will automatically apply Albumentations transforms during YOLOv5 training if albumentations>=1.0.3 is installed in your environment. See https://github.com/ultralytics/yolov5/pull/3882 for full details.

Example train_batch0.jpg on COCO128 dataset with Blur, MedianBlur and ToGray. See the YOLOv5 Notebooks to reproduce:

Good luck 🍀 and let us know if you have any other questions!

ultralytics / yolov5