EECS 274 Computer Vision Object detection Human detection

EECS 274 Computer Vision Object detection

Human detection • • HOG features Cue integration Ensemble of classifiers ROC curve • Reading: Assigned papers

Human detection with HOG • Histogram of oriented gradients • Using local gradients to represent positive and negative examples

Histogram of oriented gradients

HOG descriptors

Results with MIT dataset

Results with INRIA dataset

Parameter sweeping

Block/cell size

Results

Observations • No gradient smoothing with [-1, 0, 1] derivative filter • Use gradient magnitude (no thresholding) • Orientation voting into fine bins • Spatial voting into coarser bins • Strong local normalization • Overlapping normalization blocks

Cal Tech Pedestrian Dataset A large annoated dataset with performance evaluation

Performance evaluation

Results (cont’d)

Results (cont’d)

Results (cont’d)

Results (cont’d)

Summary • • HOG, Multi. Ftr, Ftr. Mine outperform others VJ and Shaplet perform poorly Lat. Svm trained on PASCAL dataset HOG poerforms best on near, unoccluded pedestrians • Multi. Ftr ties or outperforms HOG on difficult cases • Much room for imporvment

Daimler dataset • Recent survey in PAMI 09 • Observation – HOG/lin. SVM at higher image resolution performs well, with lower processing speed) – Wavelet-based Adaboost cascade at lower image resolution performs well, with higher processing speed

Neural network with receptive fields

Results

Cue integration Multi-cue pedestrian detection and tracking from a moving vehicle, IJCV 06

Classifier ensemble • Cascade of boosted classifiers • Variable-size blocks: 12 x 12, 64 x 128, etc. 5031 blocks in 64 x 128 image patch Fast human detection using a cascade of histograms of oriented gradients, CVPR 06

Classifier ensemble An HOG-LBP Human Detector with Partial Occlusion Handling, ICCV 09

Convert holistic classifier to local-classifier ensemble ? An HOG-LBP Human Detector with Partial Occlusion Handling, ICCV 09