-1

I have a ready-made dataset of images, where the object for recognition is already cut off along the boundaries of this object and there is no need to select it with a rectangular marker in any markup program (eg. LabelImg). Surely there is a program code or a program that can make YOLO or PascalVOC files from such a dataset of prepared images. I ask for help in finding ...

Here I found an answer that this is not possible for YOLO (without background), but it is possible that this problem can be solved for PascalVOC (I want to use with Tensorflow Object Detection): YOLO Annotation Files for Already-Cropped Images

stackuser
  • 61
  • 4

1 Answers1

1

First of all, it seems like what you're trying to do is image classification (task of assigning a class/category/label to an image).

For YOLO, it does two things: localization (locating the presence of an object in the image and draws a bounding box around it) then classification on the object inside the bounding box. This combination is called object detection (see the image below)

enter image description here

For object detection tasks, it's important to represent the target object in an uncropped image in the training dataset. This allows YOLO to learn which is the background and which is/are the target object(s).

In your case, your dataset is already cropped. YOLO will not be able to learn how to distinguish your "object" from "background" if your test set is uncropped

However, if your test set is also cropped, I recommend using image classification models instead of YOLO.

Addie Ira B. Parico
  • 607
  • 4
  • 25
  • Thank you, Parico. That is my question: "In what way can I make now from the already ready cropped dataset in any way and framework to detect these objects on the photo/video? Now my TF2 classifier (new_mobile_model.tflite) works fine with this dataset, but so far I haven't found any examples how I can use it for Object Detection - all found examples are implemented by labeling images with programs like LabelImg, which creates double work - creation of a cropped dataset for Classification and then another one for Object Detection. – stackuser Feb 03 '21 at 11:46
  • In my task with a very large dataset, I would like to use YOLO for Localization+ObjectDetection and pass the detected image for accurate classification to TF2. What is the right algorithm(frameworks, sequence...) to solve this problem? – stackuser Feb 03 '21 at 11:46