Multi Object Detection with YOLO on Artwork Dataset
Report_Yihui.pdfis my Report
codeFolder include my codes
resourceFolder are some references
Below are just some notes
- classification scheme
- cross validation
- precision vs recall
- lessons learned
- Spatial pyramid pooling
- Region proposal
If external diff > minimal internal, different graphs.
Then run greedy based on edge weight 3 times for R,G,B.
Gradually merge pixels.
- Nearest neghbor
- 8 neighbors around one pixel
- sliding window
bounding box regression
What is backgound windows?
use the pool5 features to compute a new bounding box?
Make bbox bigger towards the ground truth.
- latent SVM
pool5 is better than fc6 without fine-tuning
fc sort of stands for domain specified knowledges
- mean average precision
- detection analysis tool
- using first layer
- single out a feature and figure out all max response pictures
- spatial pyramid
RCNN main steps
- Generate regions using selective search
- Extract features on each region
- classification on features
- one class bounding box regression on features
- non-maximum suppression http://videolectures.net/iccv2015_girshick_fast_r_cnn/