Different results with same model yoloV5s

[mgabell]

Hi,
I have used your repo for Yolov5 on Jetson Orin. I created a program around your repo and basically just added a consumer/producer pattern and XML output.

Then I re-downloaded your repo and ran inference on two images using the test program.
import torch
#Model
#model = torch.hub.load('Models/YoloV5', 'custom') # or yolov5n - yolov5x6, custom
model = torch.hub.load('.', 'custom', 'best.pt', source='local')
#Images
#img = 'https://ultralytics.com/images/zidane.jpg' # or file, Path, PIL, OpenCV, numpy, list
img = '/mnt/MyBook/Images/TestRun221129_Selection/ID0_AI_Image_104_1000508050.jpg' # or file, Path, PIL, OpenCV, numpy, list
#Inference
results = model(img)
#Results
results.show() #print() # or .show(), .save(), .crop(), .pandas(), etc.

So I use same model as in "my" first program as in the latest downloaded repo. The result is different, in the "old" version I get a detection on both images, but on the latter I only get one hit.

Why is that? The same parameters is used, I did the same thing with the detect.py and
all parameters should be the same, such as

conf_thres=0.25, # confidence threshold
iou_thres=0.45, # NMS IOU threshold
max_det=1000, # maximum detections per image

Any ideas? I can try to post the code here too. I need to clean it first :-|

I saw this, and will check it out but I use Pillow.Image to read the image..
#7627

Your code to read image:
else:
Read image
self.count += 1
im0 = cv2.imread(path) # BGR
assert im0 is not None, f'Image Not Found {path}'

Whereas I

Image.open(fname)

Any pointers to where to start looking would be appreciated. I am sorry for the formatting, its just not going my way..

[glenn-jocher]

👋 hi, thanks for letting us know about this possible problem with YOLOv5 🚀. We've created a few short guidelines below to help users provide what we need in order to start investigating a possible problem.

How to create a Minimal, Reproducible Example

When asking a question, people will be better able to provide help if you provide code that they can easily understand and use to reproduce the problem. This is referred to by community members as creating a minimum reproducible example. Your code that reproduces the problem should be:

• ✅ Minimal – Use as little code as possible to produce the problem
• ✅ Complete – Provide all parts someone else needs to reproduce the problem
• ✅ Reproducible – Test the code you're about to provide to make sure it reproduces the problem

For Ultralytics to provide assistance your code should also be:

• ✅ Current – Verify that your code is up-to-date with GitHub master, and if necessary git pull or git clone a new copy to ensure your problem has not already been solved in master.
• ✅ Unmodified – Your problem must be reproducible using official YOLOv5 code without changes. Ultralytics does not provide support for custom code ⚠️.

If you believe your problem meets all the above criteria, please close this issue and raise a new one using the 🐛 Bug Report template with a minimum reproducible example to help us better understand and diagnose your problem.

Thank you! 😃

[glenn-jocher]

@mgabell yes all images should be saved to the same directory, i.e. runs/detect/exp

[mgabell]

But that does not make sense. Why would I want a copy of all source files in the output folder? I want the files with detected objects?
I made changes myself to do that, but full copy was a bit strange in my opinion.

Just to take the opportunity to give some creds! Awsome job btw! thanks for this work. Really making AI obj. detection available to the masses..

[glenn-jocher]

@mgabell if you want to customize inference I'd suggest building your own workflows using python.

YOLOv5 🚀 PyTorch Hub models allow for simple model loading and inference in a pure python environment without using detect.py.

Simple Inference Example

This example loads a pretrained YOLOv5s model from PyTorch Hub as model and passes an image for inference. 'yolov5s' is the YOLOv5 'small' model. For details on all available models please see the README. Custom models can also be loaded, including custom trained PyTorch models and their exported variants, i.e. ONNX, TensorRT, TensorFlow, OpenVINO YOLOv5 models.

import torch

# Model
model = torch.hub.load('ultralytics/yolov5', 'yolov5s')  # yolov5n - yolov5x6 official model
#                                            'custom', 'path/to/best.pt')  # custom model

# Images
im = 'https://ultralytics.com/images/zidane.jpg'  # or file, Path, URL, PIL, OpenCV, numpy, list

# Inference
results = model(im)

# Results
results.print()  # or .show(), .save(), .crop(), .pandas(), etc.
results.xyxy[0]  # im predictions (tensor)

results.pandas().xyxy[0]  # im predictions (pandas)
#      xmin    ymin    xmax   ymax  confidence  class    name
# 0  749.50   43.50  1148.0  704.5    0.874023      0  person
# 2  114.75  195.75  1095.0  708.0    0.624512      0  person
# 3  986.00  304.00  1028.0  420.0    0.286865     27     tie

results.pandas().xyxy[0].value_counts('name')  # class counts (pandas)
# person    2
# tie       1

See YOLOv5 PyTorch Hub Tutorial for details.

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

[mgabell]

This is what I did, from a bit earlier version. My miss-understanding is the input size for the analysis in detect.py.
My image is 1280x720 and you specify the inference size [640x640]. As I read here:
#2234
and here:
ultralytics/yolov3#232
The image will be resized to meet minimum stride multiple of 32.

My apparent issue was that have 640 as image inference size for my image of 1280 seems to crop the image.

As seen in the image, the black spot was detected when setting inference size to 640, but the clutter at the bottom was not.
Then I change the inference size to 1280 as the actual image size is then the clutter was found.

I will not claim scaling or cropping. Perhaps the re-scaling of the image made it impossible to detect. But if I have understood the inference size correctly I should get the full image analyzed with a 640 size but then its a matter of increasing the size incrementally until I get a suitable size for my particular use case

[glenn-jocher]

@mgabell we don't crop anything, images are resized and letterboxed (padded) to --imgsz.