levinj / pedestrian-detection-project

Pedestrian Detection Project Codes and Documentations

Pedestrian-Detection-Project

#Project Video and Paper: Project Video

Project Paper

#Project Goal:

The goal of this project is to explore how better feature representation and various visual cues can be utilized to improve detection quality.

Specifically, this project targets the fascinating and meaningful real world problem "pedestrian detection" as a test case. Using current state of the art pedestrian detector "SquaresChnFtrs" as a baseline, I leverage two approaches to increase detection accuracy. Expand 10 HOG+LUV channels into 20 channels by using DCT (discrete cosine transform); Encode the optical flow using SDt features (image difference between current frame T and coarsely aligned T-4 and T-8).

Note that this project is largely to reproduce observations/discovery in “Benenson etc., 2014 ECCV” paper. The DCT method is expected to yield 3.53% miss rate improvement, and the optical flow method is expected to yield 4.47% improvement.

#What Has Been Done:

The project started in mid-November 2014, up to now, below is achieved::

1. Got the baseline detector up and running
2. Got baseline miss rate
3. Implemented the new baseline + DCT pedestrian detector.
4. Cross verified that DCT algorithm CUDA implementation in the new detector is correct. Codes here

#What's Next:

Implement the baseline + optical flow

#Current Major Issues: Refer here for a complete list of issues and corresponding updates in this project.

pedestrian-detection-project's Issues

Big miss rate gap between two different training process configurations after adding DCT channels

After adding DCT channels into the detector, a big gap of miss rate is observed between two types of training configurations configurations:

Configuration 1:

HOG-Like features, namely 8X8 pixel block, positioned regularly each 4 pixels. Miss rate: 0.256321

Configuration 2:
HOG-multiScale, namely squares of all sizes, positioned regularly each 4 pixels. Miss rate: 0.418342

The problem can be reproduced from below version of the codes:

1b6169e

The two configuration files are

https://github.com/LevinJ/Pedestrian-Detection-Project/blob/checkdct/src/applications/boosted_learning/inria_speedy_training_config.ini
and
https://github.com/LevinJ/Pedestrian-Detection-Project/blob/checkdct/src/applications/boosted_learning/eccvw2014_squareschnftrs_inria_training.config.ini

the newly implemented baseline + DCT does not improve miss rate as expected

The detailed issue description is as below:

1. Baseline detector performace on Inria dataset is 20%
2. Have the input data decorrelated by adding DCT channels is supposed to reduce miss rate by about 3%, which is about 17%.
3. But the actual resut obtained from baseline+DCT is 42%.

Need to try performing DCT separately on each channel image

Currently DCT transformation is performed on the big image (10 channels being arranged vertically). When the image height is not of miltiples of 8 pixels, some dct block would actually cross two feature images, which might become an welcome noise in the feature pool.

Need to try performing DCT after feature shrinking, as opposed to before the feature shrinking

For the SquareChnFtrs pedestrian detector baseline this project is using, it will shrink the features in the 10 channels by a facotr of 4 in order to speed up the detection and training process.

Currently DCT is perfromed before the shrinkig, this raises the possibility that the shrining might cause DCT data lost its being "decorrelated", which is, in the first place why we want to add dct channel data. So it's worth trying put the dct transformatiion after the shrinking.

DataSequence.hpp

In StixelWorlApplication.cpp, the "helpers/data/DataSequence.hpp" is included. But I can't find this file under helpers folder. Is that a bug or do I have to make DataSequence.hpp myself?

Need to add more traces for adaboost training process to help figure out why dct method does not work as expected

Specifically. For each iteration in adaboost training:

1. which feature is selected as the weak classifier
2. the weighted sample error rate for the weak classifier
3. the updated sample weight
4. the overall error rate for the training samples
5. the margin changes:the
   difference between the weighted fraction of the weak classifiers predicting the cor-
   rect label and the weighted fraction predicting the incorrect label

and at the end of the training, for all the features selected as weak classifier:

1. which channel they are from
   2 their position
2. their weights in the final strong classifier