I.GitHub project
sg project: https://github.com/danfeiX/scene-graph-TF-release
sg dataset prepare: https://github.com/danfeiX/scene-graph-TF-release/tree/master/data_tools
py-faster-rcnn: https://github.com/rbgirshick/py-faster-rcnn
II.配置記錄 configuration records:
python版本：2.7
tensorflow版本：0.11.0
cuda版本：8.0
cudnn版本：6.0.21
III.Visual Genmoe數據集與scene-graph項目
1.VGdataset數據格式：
https://blog.csdn.net/zziahgf/article/details/79097839
注：object_id, relationship_id等爲某一類的id，並非單個物體，例如所有"man"的id均爲1058532，而不是每個人不同。
objects.json的數據格式的例子，換行表示
[
{"image_id": 1,
"objects":
[{"synsets": ["tree.n.01"],
"h": 557,
"object_id": 1058549,
"merged_object_ids": [],
"names": ["trees"]
"w": 799,
"x":0,
"y":0},...]
...}
{"image_id:" 2...
}...], "image_url": ""}
]
注，boudingbox的x、y應指其左上角座標
2.項目所需數據：
image，格式jpg
image metadata
vg scene graph其中包括object和relationships，格式均爲json
3.relationships.json格式
下面爲VG數據集中的一個圖片，其image_id爲1
所有圖片的relationships的集合，即relationships.json的格式(例子)爲：
[
{"relationships":
[{relationship描述1}, {relationship描述2}...],
"image_id": 1}, ...
]
而 relationship描述 的格式(例子)爲：
{
"predicate": "wears",
"object": {
"name": "sneakers",
"h": 26,
"object_id": 5048,
"synsets": [gym_shoe.n.01"],
"w": 52,
"y": 489,
"x": 243 },
"relationshp_id": 15937,
"synsets": ["wear.v.01"],
"subject": {
"name": "man",
"h": 262,
"object_id": 1058529,
"synsets": ["man.n.01"],
"w": 60,
"y": 249,
"x": 238 },
}
3.sh文件所需文件：
create_imdb.sh：輸入：jpg圖片及其metadata; 輸出：圖片h5格式的文件，imdb_1024.h5
create_rodib.sh：輸入
VI scene-graph-master代碼說明
核心代碼：
位置：networks/model.py
包含message pooling等核心代碼
各種網絡模型的基類：networks/networks.py