suppose I have trained a yolo for balloon object detection task. After training, during testing time what happens if I pass an image which contains no balloon? will it show the bounding box on the Image?
Yes, false negatives can also occur in YOLO.
Image conditions, i.e Indoor, Outdoor, Daytime, Night time, Background, Confusing objects(like a similar looking ball) etc. can cause this.
For reference, you can take a look at this paper : Large-Scale Object Detection of Images from Network Cameras in Variable Ambient Lighting Conditions
If you want to improve your model, train your model with no class and no label text files as well.
Follow this link
Please correct if the following is wrong or right: during test time I give an input image, yolo will divide the image in grids, the grid will be passed through a cnn and will give a feature vector which will be used to get the output vector which contains c which is probability that the grid contains object, bounding box coordinates and class probabilities. So when I pass a blank image which does not contain balloon, then c will be predicted as 0 for all the grids and no bounding box is given.
You’re right, but this is the case when the complete model is trained perfectly. As shown in the above paper, though False Negatives are fairly low (around 1% only), but this depends on how well our model has been trained, what was the quality of input data, was the lighning/background etc. different in training and inference data. There can be countless reasons for this. So, model giving c=0 in each of the cases becomes difficult sometimes.