About the choice of data set image size

[iscyy]

Search before asking

• I have searched the YOLOv5 issues and discussions and found no similar questions.

Question

Hello author, may I ask how big the photos in the data set are most suitable for the yolo model framework? For example, 640? 1280? 1920? ...For small targets, is it the picture and the larger the resolution, the better

Additional

Can the author share his experience

[github-actions]

👋 Hello @Him-wen, 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://ultralytics.com or email Glenn Jocher at glenn.jocher@ultralytics.com.

Requirements

Python>=3.6.0 with all requirements.txt installed including PyTorch>=1.7. To get started:

$ git clone https://github.com/ultralytics/yolov5
$ cd yolov5
$ 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), validation (val.py), inference (detect.py) and export (export.py) on MacOS, Windows, and Ubuntu every 24 hours and on every commit.

[glenn-jocher]

@Him-wen 👋 Hello! Thanks for asking about improving YOLOv5 🚀 training results.

Larger image sizes usually lead to better results, but take longer to process, so it's up to the user to find the right compromise between speed and accuracy for their particular application. Default image size for YOLOv5 P5 models is 640, default image size for YOLOv5 P6 models is 1280.

Most of the time good results can be obtained with no changes to the models or training settings, provided your dataset is sufficiently large and well labelled. If at first you don't get good results, there are steps you might be able to take to improve, but we always recommend users first train with all default settings before considering any changes. This helps establish a performance baseline and spot areas for improvement.

If you have questions about your training results we recommend you provide the maximum amount of information possible if you expect a helpful response, including results plots (train losses, val losses, P, R, mAP), PR curve, confusion matrix, training mosaics, test results and dataset statistics images such as labels.png. All of these are located in your project/name directory, typically yolov5/runs/train/exp.

We've put together a full guide for users looking to get the best results on their YOLOv5 trainings below.

Dataset

• Images per class. ≥ 1500 images per class recommended
• Instances per class. ≥ 10000 instances (labeled objects) per class recommended
• Image variety. Must be representative of deployed environment. For real-world use cases we recommend images from different times of day, different seasons, different weather, different lighting, different angles, different sources (scraped online, collected locally, different cameras) etc.
• Label consistency. All instances of all classes in all images must be labelled. Partial labelling will not work.
• Label accuracy. Labels must closely enclose each object. No space should exist between an object and it's bounding box. No objects should be missing a label.
• Background images. Background images are images with no objects that are added to a dataset to reduce False Positives (FP). We recommend about 0-10% background images to help reduce FPs (COCO has 1000 background images for reference, 1% of the total). No labels are required for background images.

Model Selection

Larger models like YOLOv5x and YOLOv5x6 will produce better results in nearly all cases, but have more parameters, require more CUDA memory to train, and are slower to run. For mobile deployments we recommend YOLOv5s/m, for cloud deployments we recommend YOLOv5l/x. See our README table for a full comparison of all models.

• Start from Pretrained weights. Recommended for small to medium sized datasets (i.e. VOC, VisDrone, GlobalWheat). Pass the name of the model to the --weights argument. Models download automatically from the latest YOLOv5 release.

python train.py --data custom.yaml --weights yolov5s.pt
                                             yolov5m.pt
                                             yolov5l.pt
                                             yolov5x.pt
                                             custom_pretrained.pt

• Start from Scratch. Recommended for large datasets (i.e. COCO, Objects365, OIv6). Pass the model architecture yaml you are interested in, along with an empty --weights '' argument:

python train.py --data custom.yaml --weights '' --cfg yolov5s.yaml
                                                      yolov5m.yaml
                                                      yolov5l.yaml
                                                      yolov5x.yaml

Training Settings

Before modifying anything, first train with default settings to establish a performance baseline. A full list of train.py settings can be found in the train.py argparser.

• Epochs. Start with 300 epochs. If this overfits early then you can reduce epochs. If overfitting does not occur after 300 epochs, train longer, i.e. 600, 1200 etc epochs.
• Image size. COCO trains at native resolution of --img 640, though due to the high amount of small objects in the dataset it can benefit from training at higher resolutions such as --img 1280. If there are many small objects then custom datasets will benefit from training at native or higher resolution. Best inference results are obtained at the same --img as the training was run at, i.e. if you train at --img 1280 you should also test and detect at --img 1280.
• Batch size. Use the largest --batch-size that your hardware allows for. Small batch sizes produce poor batchnorm statistics and should be avoided.
• Hyperparameters. Default hyperparameters are in hyp.scratch.yaml. We recommend you train with default hyperparameters first before thinking of modifying any. In general, increasing augmentation hyperparameters will reduce and delay overfitting, allowing for longer trainings and higher final mAP. Reduction in loss component gain hyperparameters like hyp['obj'] will help reduce overfitting in those specific loss components. For an automated method of optimizing these hyperparameters, see our Hyperparameter Evolution Tutorial.

Further Reading

If you'd like to know more a good place to start is Karpathy's 'Recipe for Training Neural Networks', which has great ideas for training that apply broadly across all ML domains:
http://karpathy.github.io/2019/04/25/recipe/](https://github.com/ultralytics/yolov5/issues/2844#issuecomment-851338384)

[iscyy]

Thanks for your answer

[drmuhammadmohsin]

But what if we have image size of 3264x2464 high resolution then again yolov5 model resize the image to the default setting or will process at the same level during training if we have no time and processing constraints ?

[glenn-jocher]

@engineermohsin if you have high resolution images of 3264x2464, YOLOv5 model will resize the images to the default setting during training unless you specify a different image size using the --img argument. Keeping the original high resolution might significantly increase processing time and resource requirements, while potentially adding minimal benefits. It's recommended to find a balance between processing constraints and model performance to achieve efficient detection results.

[atmadeep]

Thanks a lot @glenn-jocher for the insight. Can you please tell me what would be a good resolution for input training images since I have 5k (5120 x 2880) resolution videos available (taken from a go-pro). The camera also generates video_number.LRV files which are basically low resolution version with frame size (848x480). I think this is too low as the videos are really grainy and the object in far distance loses quite a bit representation.

[glenn-jocher]

@atmadeep, for your high-resolution 5k (5120 x 2880) videos, a good starting point for input training image resolution could be in the range of 640 to 1280 pixels on the long edge. This range is a balance between maintaining sufficient detail for object detection and managing computational resources effectively.

Given the original high resolution of your videos, starting at 1280 pixels might be more appropriate, as it would better preserve the details of distant objects, which could be crucial for your application. However, it's essential to consider the computational cost, as higher resolutions significantly increase processing time and memory requirements.

The low-resolution versions (848x480) you mentioned might indeed be too low for detecting small or distant objects effectively. It's worth experimenting with different resolutions to find the optimal balance for your specific use case. Remember, the best resolution depends on the nature of the objects you're trying to detect and the computational resources at your disposal.

[navidTerraNova]

Hello
I have a question regarding the input size of Yolov8 or in general the input to Yolo models, what if my images are (320 x 320), Will yolo resize the image to 640 x 640? and if it does what technique does it use for upsampling? Thank you.

[glenn-jocher]

@navidTerraNova hello! Yes, if your images are smaller like (320 x 320), YOLO models, including YOLOv8, will resize them to the model's default input size, such as 640 x 640, to ensure consistency. This resizing uses bilinear interpolation for upsampling, a common technique that's effective at preserving the image's general appearance. You can adjust the input size with the --img-size flag during training or inference if needed. Keep in mind, matching the training and inference sizes can help with performance! 😊

[iwanttograduatepls]

Hi there @glenn-jocher. I have question. Me and my partner are annotating the images and we are proceeding to training. We used the Roboflow to annotate and all our images are random image ratio which we got from YouTube and Video Recording (live). Our focus is to object detect the vehicles. Since we are assigned to be custom dataset, and we are now proceed to training using google collab. We will use the YOLOv7 (first one) as the model and we confused on which -img will use. The Median Image Ratio based on Roboflow was 1920x1080. Average Image Size was 2.07mp.

1. Is it good to use the 640 or 1280?

[glenn-jocher]

Hello @iwanttograduatepls! Thanks for reaching out with your question.

When deciding on the image size (--img) for training your custom dataset, it's important to balance between computational efficiency and the level of detail required for accurate object detection. Here are some considerations:

1. Default Settings: Starting with the default image size of 640 is generally a good practice. This helps establish a performance baseline and allows you to identify any potential issues early on.

2. Higher Resolution: If your dataset contains many small objects or if the objects are far away, training at a higher resolution like 1280 can help improve detection accuracy. However, this will require more computational resources and longer training times.

3. Consistency: Ensure that the image size used during training matches the size used during inference for optimal performance.

Given that your median image ratio is 1920x1080 and the average image size is 2.07MP, starting with 640 and then experimenting with 1280 could be a good approach. Here’s how you can set the image size in your training command:

python train.py --data custom.yaml --weights yolov7.pt --img 640

If you find that the model benefits from higher resolution without significant computational drawbacks, you can switch to:

python train.py --data custom.yaml --weights yolov7.pt --img 1280

Remember to monitor your training results, including metrics like precision, recall, and mAP, to determine the impact of different image sizes.

Good luck with your training, and feel free to share your results or ask further questions! 😊

[drmuhammadmohsin]

Thank you for your response. I tried the progressive approach, starting from the default size of 640 and then increasing the resolution to 1280 and 1920. With higher resolution, the results improved in my case for detecting small electronic components on PCBs, but at the cost of higher memory usage and longer training times. The trained model size is very large in this case, making it very challenging to perform inference on an edge device (Jetson Nano). Do you have any suggestions, especially if we divide the image into a grid of 640x640 and train the model on each grid cell? What do you think? Thanks again! Best regards, Mohsin

…

________________ Muhammad Mohsin PhD Researcher in Computer Science DIBRIS - Università degli Studi di Genova - Italy Email: ***@***.*** Mob. +39 345 5017833

On Thu, Nov 16, 2023 at 1:55 AM Glenn Jocher ***@***.***> wrote: @engineermohsin <https://github.com/engineermohsin> if you have high resolution images of 3264x2464, YOLOv5 model will resize the images to the default setting during training unless you specify a different image size using the --img argument. Keeping the original high resolution might significantly increase processing time and resource requirements, while potentially adding minimal benefits. It's recommended to find a balance between processing constraints and model performance to achieve efficient detection results. — Reply to this email directly, view it on GitHub <#5851 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AJ7VLLKH77EBRTTTU63M76DYET6ZRAVCNFSM5JGI77AKU5DIOJSWCZC7NNSXTN2JONZXKZKDN5WW2ZLOOQ5TCOBRGI4TSOJYHE4A> . You are receiving this because you were mentioned.Message ID: ***@***.***>