Pierre Sermanet
PhD in deep learning for vision, speech and robotics
News | Videos | Software | Data | Publications | Citations | LAGR | Eurobot | Resume

Preprocessed SVHN

  • The Non-commercial only SVHN dataset is reproduced here for convenience only. Full copyright remains with the authors (Yuval Netzer et al.). Please cite their paper if using their dataset, see http://ufldl.stanford.edu/housenumbers/

  • This dataset is available in different formats:

  • This dataset is organized as follow:
    • Training set: 598388 samples
    • Validation set: 6000 samples
    • Testing set: 26032 samples

  • The images, labels and names of each class each have their own matrix. For example the validation set is composed of the following matrices:
    • svhn_ynuv7_val_data.mat: 6000x3x32x32
    • svhn_ynuv7_val_labels.mat: 6000
    • svhn_ynuv7_val_classes.mat: 10x2

  • The original RGB 32x32 inputs were converted to YUV inputs where the Y channel is contrast normalized (see examples below).

Local-contrast normalized Y channel:

U channel:

V channel:

Inria Test + Ignored true positive bounding boxes

The original INRIA pedestrian Test groundtruth is problematic as not all pedestrians are labeled (mostly harder ones such as people behind glass doors, people reflections, small pedestrians or even ones that are not easily noticable by a human eye.

Algorithms that manage to detect these unlabeled pedestrians have a decreased performance with the original groundtruth.

To report the true performance, you can use the corrected groundtruth below which includes all the unlabeled pedestrians with the "ignore" label which means these detections will not count as errors.

  • inria_test_ignore.vbb (in Piotr Dollar's pedestrian evaluation Matlab format, goes into "data-INRIA/annotations/set01" path)

ConvNet bounding boxes for pedestrian detection

These bounding boxes were produced by the model described in our CVPR'13 paper "Pedestrian Detection with Unsupervised Multi-Stage Feature Learning" and its supplement:
Below is a modified version of Piotr Dollar's benchmarking code (v3.0.1) available here. This modified code was used to generate plots and tables in the paper above. Full rights remain to Piotr Dollar. The modifications include: a switch to use full or partial AUC, automatic generation of a latex table of all results and some minor cosmetic tweaks.

Last update: March 27th, 2013