Different conffidence results between ONNX and Pt, issue with the letterbox? about ultralytics

[github-actions]

👋 Hello @yenke21, thank you for your interest in Ultralytics YOLOv8 🚀! We recommend a visit to the Docs for new users where you can find many Python and CLI usage examples and where many of the most common questions may already be answered.

If this is a 🐛 Bug Report, please provide a minimum reproducible example to help us debug it.

If this is a custom training ❓ Question, please provide as much information as possible, including dataset image examples and training logs, and verify you are following our Tips for Best Training Results.

Join the vibrant Ultralytics Discord 🎧 community for real-time conversations and collaborations. This platform offers a perfect space to inquire, showcase your work, and connect with fellow Ultralytics users.

Install

Pip install the ultralytics package including all requirements in a Python>=3.8 environment with PyTorch>=1.8.

pip install ultralytics

Environments

YOLOv8 may be run in any of the following up-to-date verified environments (with all dependencies including CUDA/CUDNN, Python and PyTorch preinstalled):

• 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 Ultralytics CI tests are currently passing. CI tests verify correct operation of all YOLOv8 Modes and Tasks on macOS, Windows, and Ubuntu every 24 hours and on every commit.

[glenn-jocher]

@yenke21 hey there 👋! It appears there's a slight discrepancy in confidence scores between the ONNX and PyTorch (.pt) models, possibly linked to the pre-processing step with letterboxing. The purpose of the same_shapes check within the predictor.py is to determine if all images within your batch have the same shape. If true, and if you're using a PyTorch model (.pt), the LetterBox process is auto-adjusted for each image based on the model's stride, optimizing the padding and scaling process.

From your results, the confidence scores align closely between PT_WITH_LETTERBOX and ONNX, suggesting that using the letterbox transformation aligns the input more closely with how the ONNX model expects it compared to raw input or list input. This discrepancy can arise due to differences in how ONNX and PyTorch handle image inputs—ONNX might be inherently applying similar pre-processing as the LetterBox class, leading to matching results when the same pre-processing is explicitly applied to the .pt model inputs.

However, considering the small variation in confidence, it’s also wise to review other factors that might be impacting the ONNX inference. Differences in the ONNX and PyTorch runtime environments, such as precision handling (float32 vs. float16), could also contribute to slight discrepancies.

If consistent input processing is crucial for your application, you might consider explicitly applying the same pre-processing (like LetterBox) across both model formats to ensure uniformity in inference conditions. Additionally, verifying the inference environment's precision settings could further align the outputs.

Hope this sheds some light on the situation! If the issue persists, diving deeper into pre-processing steps and runtime environment configurations might reveal more insights. 😊🔍