Description

This is Pedestrian Detection with Deep Compression developed by
Kevin Le [email protected] , Jeanette Nguyen [email protected] , Thomas An [email protected] , Christian Gunther [email protected]

Requirements

requires Python v3

requires PyTorch >=0.4

install cupy:
$ pip install cupy-cuda80

install other dependencies:
$ pip install -r requirements.txt

Optional, but strongly recommended: build cython code nms_gpu_post:

$ cd model/utils/nms/
$ python build.py build_ext --inplace

Code Organization

demo.ipynb - Run a demo of our code
Train.ipynb - Run the training of our model

core/  
-- data/  
|  -- __init__.py - Faster RCNN data init  
|  -- d_util.py -  Faster RCNN data utils
|  -- dataloader.py - Faster RCNN dataloader
-- __init__.py - Faster RCNN init
-- logger.py - Faster RCNN logger for data and model

data/
-- __init__.py - data init
-- __caltech_dataset.py - Split the data into different sets
-- dataset.py - Preprocess data and get it in the necessary format
-- util.py - Utility functions to preprocess the data

dataset/
-- data_test.csv - Caltech Dataset Test sets 06-10
-- data_train.csv - Caltech Dataset Train sets 00-05
-- data_val.csv - Caltech Dataset Validation part of sets 00-10

misc/
-- convert_caffe_pretrain.py - Convert caffe pretrained weights to be usable by our model

model/
-- compression/
|  -- PruningClasses.py - Module to prune weights from the model
|  -- __init__.py - model compression init
|  -- prune_utils.py - Utils for pruning
|  -- quantization.py - Function to quantize the weights
|  -- vgg16.py - Modified VGG Network to support pruning and quantization of weights
-- model_deprecated/
|  -- FasterRCNN.py - Unused model
|  -- RPN.py - Unused region propasal network
|  -- VGG16.py - Unused model
-- utils/
|  -- nms/
|  |  -- build/temp.linux-x86_64-3.6/
|  |  |  -- _nms_gpu_post.o
|  |  -- __init__.py - Faster RCNN non-maximum suppression init
|  |  -- _nms_gpu_post.c - Faster RCNN nms C extension
|  |  -- _nms_gpu_post.cpython-36m-x86_64-linux-gnu.so
|  |  -- _nms_gpu_post.pyx
|  |  -- _nms_gpu_post_py.py - Faster RCNN cms code
|  |  -- build.py - Faster RCNN Build the cython code for nms
|  |  -- non_maximum_suppression.py - Faster RCNN Suppress bounding boxes according to their IoUs
|  -- __init__.py - Faster RCNN model utils init
|  -- bbox_tools.py - Generate bounding boxes and perform calculations on them
|  -- creator_tool.py - Generate proposal regions
|  -- roi_cupy.py - Faster RCNN generate regions of interest
-- utils/deprecated/
|  -- bbox.py - Unused bounding box file
|  -- config.py - Unused config file
|  -- network.py - Unused network file
|  -- proposal_layer.py - Unused proposal layer file
-- __init__.py - Faster RCNN model init
-- faster_rcnn.py - Faster RCNN model
-- faster_rcnn_vgg16.py - Faster RCNN model based on vgg16
-- region_proposal_network.py - Region Proposal Network introduced in Faster R-CNN
-- roi_module.py - Region of Interest Module

tools/
-- __init__.py - tools init
-- benchmark_model.py - Measures framerate of the evaluation
-- plot_annotations.py - Draw bounding box annotations on images
-- preparte_dataset.py - Generate data csv files
-- visualize_dataset.ipynb - Display images with bounding boxes

utils/
-- __init__.py - utils init
-- array_tool.py - Tools to convert specified type
-- config.py - Settings to configure the model 
-- constants.py - Declared constants
-- eval_tool.py - Tools to evaluate the accuracy of our detections
-- size_utils.py - Get size of our model
-- vis_tool.py - Tools to help visualize the images with bounding boxes

__init__.py - Faster RCNN init

demo.ipynb - Demo of our evaluation to detect some pedestrians

eval.py - Evaluate our model's MAP

prune.py - Trains a model with pruned weights

quantize.py - Quantizes the model weights

requirements.txt - Requirements that must be installed for the model to run

train.ipynb - Notebook to rerun the training if need be

Train.py - Run the code to train our Faster RCNN model

trainer.py - Wrapper for conveniently training

Prepare Data

To download the dataset copy the github repository: $ git clone https://github.com/mitmul/caltech-pedestrian-dataset-converter.git

To prepare the dataset, run prepare_dataset.py in the tools directory: $ python prepare_dataset.py --path=/path/to/output --data-dir=/datasets/ee285f-public/caltech_pedestrians_usa/

Output for example run: $ python prepare_dataset.py --path=../data --data-dir=/datasets/ee285f-public/caltech_pedestrians_usa/

Note, preparing the data is not necessary. We have included the our own prepared csv files (train, val, and test) in the repository as well in the dataset folder.

Pretrained Model

Download the pretrained weights from our model here.