Install SiamFC-TensorFlow on RPi

在笔记本上和 RPi 上都进行测试.

目录: chip:~/workspace/object-tracking/siamfc/siamfc-tf pi3:~/workspace/

按官方指南, 使用 virtual-env, 创建一个独立的环境, 不依赖于系统里已经安装的 python 库.

siamfc-tf 需要 tensorflow-gpu==1.1.0, tensorflow-gpu 依赖 libcublas, 会在运行时加载 libcublas 动态库, 这个库需要 CUDA.

需要把 tensorflow-gpu 换成 tensorflow, 但是直接使用 pip 安装 tensorflow 失败, 找不到 tensorflow 这个包. 于是使用 tensorflow-on-raspberry-pi 提供的二进制安装. tensorflow-1.1.0

使用 pip 安装依赖 matplotlib 时, 编译错误:

src/_png.cpp: In member function 'Py::Object _png_module::write_png(const Py::Tuple&)':
src/_png.cpp:147:51: error: 'npy_PyFile_Dup' was not declared in this scope
...

change numpy version from 1.11.1 to 1.6.2 NO. I install it manually,

OK, 可以运行.

原输入是一个文件夹, 里面都是图, 以及一个框. 输出, 每帧一个框, 图片可以显示.

改成. 输入是摄像头, (初始框暂时不管

region_to_bbox, region is either 4 dim or 8 dim. 4 dim is rect, x y w h, 8 dim is poly, … not sure yet. anyway , only 4 is used here

Where did it open the image file? tracker.py 70, it use a for loop to process each frame.

sess.run use a feed_dict, which requires a filename for a frame. I need to stop it sess is a tf.Session. what do you what ? see official doc ! It seems that the tf graph is in the global environment. sess.run(), run it.

run : Runs operations and evaluates tensors in fetches.

feed_dict argument allows the caller to override the value of tensors in the graph. See run() for more information.

so the computation has already been defined in the graph definition part.

it used a tf.train.Coordinator() to load image file!!!

A coordinator for threads. This class implements a simple mechanism to coordinate the termination of a set of threads.

修改 graph, 发现不知道 tensorflow 如何输入图片. 计算图怎么写?

看了下 Session 和 run,

使用 feed_dic 输入 graph. 弄个 place_holder 接受就行了.

这个真的.. 我真的没把这里当成博客啊.. 这不是把屎拉在路上么..