Objects Detection with YOLO on Artwork Dataset ( you only look once system)
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
code
resource
README.md
Report_Yihui.pdf
prog7.pdf

README.md

Multi Object Detection with YOLO on Artwork Dataset

  • Report_Yihui.pdf is my Report
  • code Folder include my codes
  • resource Folder are some references
Below are just some notes

writeup

  • dataset
  • features
  • partition
  • classification scheme
  • cross validation
  • precision vs recall
  • accuracy
  • lessons learned

#faster RCNN
RCNN

  • Spatial pyramid pooling
  • Region proposal

useful link

Tutorial on Tools for Efficient Object Detection

segmentation

Kmeans

mean shift

Graph

If external diff > minimal internal, different graphs.
Then run greedy based on edge weight 3 times for R,G,B.
Gradually merge pixels.

Distance measure:

  • Nearest neghbor
  • 8 neighbors around one pixel

Localization

  • HoG
  • SIFT
  • “neocognitron”
  • 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.

detection

  • latent SVM
  • convNet

Classifier

  • softmax
  • SVM
  • KNN?

ablation study

pool5 is better than fc6 without fine-tuning
fc sort of stands for domain specified knowledges

evaluation

  • mean average precision
  • detection analysis tool

visualization

  • using first layer
  • single out a feature and figure out all max response pictures

other approach

  • overfeat
  • spatial pyramid

RCNN main steps

  1. Generate regions using selective search
  2. Extract features on each region
  3. classification on features
  4. one class bounding box regression on features
  5. non-maximum suppression http://videolectures.net/iccv2015_girshick_fast_r_cnn/