How to assign the model id and model version in batch_input_params.json ?

Albert ChungAlbert Chung
in KL720 related discussion

Hello,

I want to know how to assign the model id and model version in batch_input_params.json.

Is there any reference about it ?

Thanks!

Tagged:

Comments

  • Albert ChungAlbert Chung

    你好,

    模型相關的參數如 model_id, image resolution, model input dimension, class list等相關參數

    請問我應該修改哪個文件 ?

    謝謝!

  • Otis HsiehOtis Hsieh

    Hi Albert Chung,

    For the first question, if you want to modified the json file, you can refer here :

    http://doc.kneron.com/docs/#toolchain/command_line/#b-how-to-configure-the-batch_input_paramsjson

    B. How to configure the batch_input_params.json?

    image.png

    For the second question, I use PLUS kl720_demo_generic_inference as example. You can modify the descriptor.

    image.png

    As for the class list, in our YOLO example, you can refer this: http://doc.kneron.com/docs/#plus/introduction/yolo_object_name_mapping/

    Thanks.

  • Albert ChungAlbert Chung
    edited December 2021

    Hi Otis Hsieh,

    感謝您的說明

    目前使用model id 為 KNERON_YOLOV5S_640_640_3

    跑 704x480 及 640x640影片的偵測框還算正常

    但是跑1920x1080 影片出現偵測框往下偏移的現象

    確定過_image_width 和 _image_height 和影片相同

    postprocess 在dongle處理

    是否是postprocess 的設定問題 , 該如何來修改這部分 ?


    謝謝您!

  • Otis HsiehOtis Hsieh

    Hi Albert Chung,

    這部分是因為720在inference之前

    圖片preprocess的部分會把圖片resize成跟model size相同

    所以不足的部分會進行Padding，

    以你的例子來說，1920X1080會等比例resize成640X360，然後padding上 640-360=280

    但是padding有上限255，所以會導致最後的bounding box會有些許偏移


    我們有個kl720_demo_generic_inference_crop的code，

    他可以crop圖片中的特定區域，進行inference

    image.png

    包括你要crop區域的數量，以及他們的座標 / 長寬

    看你的場景需求調整，謝謝

The discussion has been closed due to inactivity. To continue with the topic, please feel free to post a new discussion.

© Kneron, Inc.