search

HiKapok / SSD.TensorFlow

State-of-the-art Single Shot MultiBox Detector in Pure TensorFlow, QQ Group: 758790869
Apache License 2.0
323 stars 153 forks source link

Bug: Using images out of ImageSets\Main\trainval.txt for training #102

Open jamiechoi1995 opened 5 years ago

jamiechoi1995 commented 5 years ago

In convert_tfrecords.py, it converts all images in VOCdevkit\VOC2012\JPEGImages into tfrecords for training, https://github.com/HiKapok/SSD.TensorFlow/blob/b47ff6164c8925a8bbccc593719d5bbbab996058/dataset/convert_tfrecords.py#L366-L370

However, not all the images in VOC2012 are for detection, there are only 11540 images specified in ImageSets\Main\trainval.txt for detection. But in JPEGImages there are 17125 images, these extra images are for action classification, so they only annotate human, ignoring other classes. Eg, 图片 图片

Using these images will be harmful for training.

jamiechoi1995 commented 5 years ago

after removing the images and using 11540 images to train the model, I got 0.7785 mAP with the default settings using model-ckpt-116604