First semester of 2018

4th Practical Activity - Pedestrian Detection

0. Download the databases

   • INRIA Person dataset (970Mb)

     • 1208/566 positive windows (train/test) -- 64 x 128 bounding box size with 16/3 border pixels (train/test)
     • 1218/453 negatives images (train/test) -- 2592 x 1944 pixels

   • MIT (CBCL PEDESTRIAN DATABASE #1) -- 924 images (10/22 Mb) with 64 x 128 PPM format images for training

   • Caltech Pedestrian Detection Benchmark (6Gb/5Gb)

   • ETH dataset

1. Implement a code to load the databases [code/interface]

2. HoG Computation

3. Training SVM

4. Hard negative Mining

5. Sliding windows

6. Non-Maximum suppression

7. Evaluation: Miss rate vs FPPW curve

Evaluation

Send by moodle the generated codes until the begining of the seminars