VHD with non-deep models

Requirements:

The following libraries are required to run the codes.

• sklearn
• multiprocessing
• itertools
• xgboost

They each supports: machine learning tools, multi-process reading data, Loop testing to tune parameters, GBDT lib.

Run the code:

1. Configurations:

    Mannually set PATHs and hyper-params in configs.py as following:
    1,PROJECT_ROOT         the path to project
    2,BAIDU_VH_ROOT        the path to dataset
    3,MIN_FEATURE          the mini value of feature data( for computing svm kernels)
    4,METRICS              Use AP here
    5,ALTER                smoothing params
    6,GRID_SEARCH_PARAMS   grid search params

2. train the baidu dataset:

sh scripts/watershed_train.sh

3. test the baidu dataset(computing result on both val and test):

sh scripts/watershed_test.sh

Results:

Single model's performance (sample rate: 1:100 on train set):

tested on validation set

model name	AP
(1) Logistic regression	81.7%
(2) XGBoost	82.8%
(3) SVM (linear kernel)	81.5%
fusion (1) and (2)	83.2%

Tips: SVM with x2 kernel or rbf kernel may bring higher result, but we should consider more of efficiency here.

Detection performance

highlights score curves are generated from above, here we smooth them with multiple filters and then watersh them to final results.

set: IoU=0.5

filter length	AP	recall
4	0.41	0.56
8	0.49	0.51
12	0.50	0.52
20	0.45	0.51
40	0.39	0.42
80	0.31	0.38
fusion 8,12,20,40,80 + TBP	0.45	0.81
fusion 4,8,12,20,40,80 + TBP	0.43	0.85

Tips: NMS param threshold value is slightly tuned from 0.3~0.7.

Else, TBP is method: Throw false positive By Prior.