1. 1 Accurate Object Detection with Joint Classification- Regression Random Forests Presenter ByungIn Yoo CS688/WST665

2. 2 Contents ●Introduction ●Motivation ●Main Idea ●Details ●Experiments ●Conclusion

3. 3 Introduction ●Object detection is one of the most important task for objects recognition as well as images searching. ●Object Detection + Bounding-box Regression (Localization) in a sliding window approach with a single model (Random Forest). Classification: Car Regression: 1(width):0.6(height)

4. 4 ●Problems ●Low-accurate Bounding-box Localization ●Low-accurate Label Classification Motivation 89.3% overlap (Proposed method) 59.6% overlap (Previous methods) Ground-truth  How we improve performances of localization and classification simultaneously?

5. 5 Main Idea ●Joint Classification-Regression Random Forest (JCRF)  Classification: Predict object probability  Regression: Estimate bounding-box aspect ratio ●What things are novel?  More accurate object detection and localization method in a single model! Training data : Car region, Background : Aspect ratio of the region …… Tree1 Treet Obj./Back.Aspect ratio x 0.8 Obj./Back.Aspect ratio 0.7 x Random Forest (JCRF) Testing Result : Car region (x,y) : Aspect ratio of the region 0.75 (width-height) TrainingTesting

6. 6 ●Training Data = image = {height x width x 10 feature channels} Details – Object Detection Model Label = {background, object} Actual height (width is normalized to 100 pixels) Blue boxes ( h x w ) show positive training data(Object). Z i show actual height of objects. Negative data (Background)

7. 7 Details – Training JCRF (1/3) ●What is Random Forest? : Ensemble of multiple decision trees ●Split Node: Find and store a best splitting parameter. ●Leaf Node: Store class probabilities and an aspect ratio. …… Tree 1 Tree t Obj./Back.Aspect ratio x 0.8 Obj./Back.Aspect ratio 0.7 x Split Node Leaf Node Simply averaging class probabilities and aspect ratio from all trees!

8. 8 Details – Training JCRF (2/3) ●Two split node types are employed. ●Binary Classification Node ( Object or Background? ) ●Regression Node ( How long is the object height? ) ●Randomly decided which types are being optimized in an each split node. ●Classification Split nodes ●Objective: Find a good splitting rule which minimize the Entropy of the classes between left and right dataset. ●Regression Split nodes ●Objective: Find a good splitting rule which minimize the height variance of the between left and right dataset.

9. 9 Details – Training JCRF (3/3) Location1Location2ThresholdChannel Training Data What is the best parameter to split? Splitting function Pixel value 1 of channel c Pixel value 2 of channel c threshold Shannon Entropy Location1 Location2

10. 10 Details – Testing JCRF (1/2) ●For detecting objects in test images, a standard sliding window W is utilized. ●Detection score s of a given image x in a window W ●Object height z of a given image x in a window W ●The resulting detection D of a window W k is the scale of the detection x,y: location of Ww: width of Wz: height of W s: Object detection score of W F R : Regression Function of a JCRF T: Number of Trees in a JCRF F C : Classification Function of a JCRF T: Number of Trees in a JCRF

11. 11 Details – Testing JCRF (2/2) ●Early stopping is utilized to boost a testing speed. … tree 1tree T xx … tree t x Early Stopping Criterion: Testing Progress Detection threshold Current summation of scoresUpper bound of remaining scores

12. 12 Experiments (1/3) ●Evaluation Criterion: Pascal overlap  IoU (Intersection over Union) Detected RegionGround-truth :True :False

13. 13 Experiments (2/3) ●Precision-recall curve for the bounding-box accuracy.  Proposed methods shows best performance. [ TUDpedestrian Dataset ]

14. 14 Experiments (3/3) ●Tightening the Pascal Overlap Criterion  Proposed methods shows best performance. [ ETHZcars Dataset ]

15. 15 Conclusion ●Random forest based object detection and predicting aspect-ratio method is proposed. ●Joint Classification-Regression Forest exploits class labels as well as actual heights during both training and testing. ●Proposed detection model recognize more accurate object regions than related state-of-the art approaches.

16. 16 Appendix – More Experiments Saturation! Classification onlyClassification + Regression Diverse locations!  Separate Different Views