PaddlePaddle
Inference on a single image in real time
问题确认 Search before asking
- [x] 我已经搜索过问题，但是没有找到解答。I have searched the question and found no related answer.
请提出你的问题 Please ask your question
The project that I am currently working is “Real Time Detection” of casting defects on Xray Radiography images of size 1000 x 1000 and 1024 x 1024...... After a lot of deliberation on PaddleDetection configs, I chose “smrt” and currently using “ppyoloe_crn_m_300e_battery_1024.yml” as the baseline and I have trained on a custom dataset of 4 object classes. Training and Evaluation results are quite impressive and I have to deploy it “live” on the production system.
The requirement is “Single Image” detection in real time on a continuous basis (frequency of 3/4 seconds for a new image as the component/part which is held in a Industrial robot, flips it through different orientations and Maximum latency of 500 ms for inference) and a typical Xray Radiography machine, currently used in production, will generate close to 5000 images in a shift.
And currently, quality inspectors trained on reading Radiography images does the detection of defects in real time
When I use “/deploy/python/infer.py” as the code and build a UI based on PyQt5, I am not gettng the “Live Image Viewer” running, as “infer.py” is primarily written to infer on a “single image” or an “image_folder” or a “video_file” as the results are written to a single “bbox.json” file.
Can you give me some insights and based on which, if I have to share additional information, I will be do.....?
