【ECCV 2016】SiamFC 细节笔记

对SiameseFC中的一些细节做个记录

Training

对y的GT标记为$y\in \{+1,-1\}$，有logistic loss
$$\ell (y,v)=\log (1+\exp (-yv))$$
其中，v由模板分支和候选分支的到的分数值。
Score map的损失定义为：
$$L(y,v)=\frac{1}{|\mathcal{D}|}\sum _{u\in \mathcal{D}}\ell (y[u],v[u])$$
在对于每个position有lable y的情况下，随机梯度下降SGD求解卷积网络参数$\theta$
$$\arg \underset{\theta }{\min }\underset{(z,x,y)}{E}L(y,f(z,x;\theta ))$$

• 提取的图片以目标为中心，缺失部分用RGB均值填充。

• 训练输入的两帧最多相隔T帧。

在图片宽高比不变的情况下，以图片中心为半径的R区域内都被认为是正例。
$$y[u]=\{\begin{array}{ll}+1& \text{ if }k\Vert u-c\Vert \le R\\ -1& \text{ otherwise }\end{array}$$

• 以target为中心的目标训练被认为更有效

实际操作

数据集管理

数据集的操作顺序大概是

1. Signup (VID)
2. Download and unzip the full original ImageNet Video dataset (the 86 GB archive).
3. 移动验证集
4. Run `./video_ids.sh /path/to/original/ILSVRC2015` to generate `vid_id_frames.txt`.
5. Run `per_frame_annotation.m` for all 5 folders.标记frame
6. Run `parse_objects.m` for all 5 folders.标记objects
7. Run `vid_setup_data.m` to generate your own `imdb_video.mat`. 
8. Add a field `.set` which is 1 for ILSVRC15-VID training videos (folders `a`, `b`, `c` and `d`)and 2 for ILSVRC15-VID validation videos (folder `e`) (to not be confused with the validation videos used during tracking evaluation, which instead come from VOT and TempleColor). For ECCV'16 SiamFC the training/validation split was decided only inside `experiment.m` code.
9. Duplicate the tree structure of the ILSVRC15 folder (without copying the data!)
10. Run `save_crops.m` to generate crops for all the videos indexed by `imdb_video.mat` and save them on disk.
11. Run `vid_image_stats.m` and save the output in a `.mat` file. It will be used for data augmentation during training.