wikke / ssd_keras Goto Github PK

Can I add new set of images to VOC dataset and how to add labels of those newly added images

下载VOC2007和PASCAL_VOC_SSD300_weight_checkpoint-epoch-16-loss-3.3392.hdf5文件，运行SSD.py，可以识别自行车、人物，图片上都能正确标注区域，但是所有带猫的图片没有标注。

我自己fork了你的项目，https://github.com/sunjl/keras_ssd，修改了SSD.py文件：

OUTPUT_DIR = './output'
MODEL_H5_NAME = 'weight.hdf5'
# MODEL_H5_NAME = 'PASCAL_VOC_SSD300_weight_checkpoint-epoch-16-loss-3.3392.hdf5'
model.save_weights(os.path.join(OUTPUT_DIR, MODEL_H5_NAME))
model.load_weights(os.path.join(OUTPUT_DIR, MODEL_H5_NAME))

自己训练模型，无法识别pics任一图片，怎么回事？
运行环境：Ubuntu 16.04，CUDA 8.0，Python 3.5，Keras 2.0.9，Tensorflow 1.3.0
GTX960显卡，4G显存不够用，所以把SSD.py文件中的batch改为1

Hi!
I am trying to understand SSD algorithm. Your implementation is the most understandable among all I've found. But still I have some questions and I hope you can help me with it.
The first question is: why do we need variances? As I understand, it is used to normalized coordinates. But why do we need normalization? For the better performance?

Also, what is the difference between location coordinates and priors coordinates?

SSD.ipynb not opening