Domain Adaptation for Pedestrian Detection Based on Prediction Consistency

Table 1

The descriptions of the datasets for training.

Dataset	Training set
	Positives	Negatives
MIT_Traffic_DA	Source domains: total of 4833 samples, INRIA, MIT CBCL, and CAVIAR WalkbyShop1cor (118 frames); Target domain: 790 samples, MIT_Traffic (top 210 frames)	6000 samples and INRIA
CAVIAR_DA	Source domains: total of 5054 samples, INRIA, MIT CBCL, and MIT_Traffic (top 210 frames); Target domain: 569 samples and CAVIAR WalkbyShop1cor (118 frames)
MIT_Only	790 samples and MIT_Traffic (top 210 frames)
CAVIAR_Only	569 samples and CAVIAR WalkbyShop1cor (118 frames)
All	Total of 5623 samples and INRIA + MIT CBCL + MIT_Traffic (top 210 frames) + CAVIAR WalkbyShop1cor (118 frames)