0.25
0.5
0.75
1.25
1.5
1.75
2
Do We Need More Training Data or Better Models for Object Detection?
Published on Oct 09, 20124198 Views
Datasets for training object recognition systems are steadily growing in size. This paper investigates the question of whether existing detectors will continue to improve as data grows, or if models
Related categories
Chapter list
Do we need more training data or better models for object detection?00:00
Object detection00:17
Current state of object detection00:45
Current state of object recognition01:33
Scanning window classifier detection02:21
Bayes Risk03:00
Ideal Behaviour03:34
Model Bias (1)03:53
Model Bias (2)04:01
Ideal Behaviour: Complexity-‐Generalization tradeoff (1)04:25
Ideal Behaviour: Complexity-‐Generalization tradeoff (2)04:33
Experiment #105:00
Performance as training set grows (1)05:25
Performance as training set grows (2)05:40
Performance as training set grows (3)05:56
Tune the regularization parameter C06:31
Experiment #207:05
Single template face model (1)07:44
Single template face model (2)08:18
SVM is sensitive to outliers08:32
Learned templates09:08
Experiment #309:34
Supervised clustering09:44
Human‐in-the-loop clustering10:28
Bus Category11:01
PASCAL 10x Dataset11:55
10x training dataset performance12:36
Have we reached the Bayes Risk for the HOG feature space?13:42
Deformable part model (DPM)14:12
Alternate view of DPM15:15
Why does DPM outperform rigid mixtures?16:05
Experiment #5: Rigid Mixture of Parts16:46
Performance of supervised DPM17:12
State of the art face detection17:54
Lesson learned18:28
Thanks19:03