Repulsion Loss

This repository implements the code for Repulsion Loss: Detecting pedestrians in a crowd. This page(code, project and presentation) is the submission for Group B for final project for the course CS 256: Topics in Artificial Intelligence, Section 2 led by Prof. Mashhour Soul, PhD. The group members include:

• Rakesh Nagaraju
• Vidish Naik
• Joel Alvares
• Charulata Lodha

The code maybe used for educational and commercial use under no warranties. For questions on this project and code please reach out to: Rakesh Nagaraju at rakesh.nagaraju@sjsu.edu

Requirements:

• PyTorch: 0.4.0
• TorchVision: 0.2.0
• CUDA: 8.0

Preparation

Download PASCAL VOC 2007 dataset

Clone this repository

git clone https://github.com/Raki2205/Raki2205-Pedestrian-Detection-using-Faster-RCNN-with-Repulsion-Loss.git

cd into the folder

cd RepulsionLoss

Export the path for CUDA:

export CUDA_HOME="/usr/local/cuda-8.0"
export LD_LIBRARY_PATH="/usr/local/cuda-8.0/lib64":$LD_LIBRARY_PATH
export PATH="/usr/local/cuda-8.0/bin:$PATH"

Verify that CUDA is indeed set to CUDA8

nvcc -V

nvcc: NVIDIA (R) Cuda compiler driver Copyright (c) 2005-2016 NVIDIA Corporation Built on Tue_Jan_10_13:22:03_CST_2017 Cuda compilation tools, release 8.0, V8.0.61

Install the requirements

pip install -r requirements.txt

Run the startup.sh script

sh startup.sh

This will compile the required libraries in the lib folder and also download a pre-trained model in data/pretrained_model directory and download PASCAL VOC dataset in data folder(if not just download Pascal VOC 2007 dataset from the internet and put in the folder named data).

NOTE: The libraries in the current state are compiled for Nvidia Tesla K80 GPU. Change the CUDA_ARCH argument in lib/make.sh if you're using a different GPU. Refer this link for the correct architecture for your GPU. You can also refer the CUDA WIKI to get the correct architecture.

Running the model

python trainval_net.py --dataset pascal_voc --net vgg16 --bs 4 --nw 1 --cuda

• Keep batch size (bs) < 4 on a Tesla K80 GPU as it will run out of memory
• Currently only works for vgg16. Future updates might include resnet101.
• dataset can be changed to pascal_voc_0712 for PASCAL VOC 2012

Model download

For testing the model download the pre-trained weights and store it in models/vgg16/pascal_voc/faster_rcnn_1_10_2504.pth

Testing

python test_net.py --net vgg16 --checksession 1 --checkepoch 10 --checkpoint 2504 --cuda

Demo

Store the images that you need to run in the images folder. New images with detections will be stored in the same folder

python demo.py --cuda --load_dir models --net vgg16 --checksession 1 --checkepoch 10 --checkpoint 2504

Credits

This project was conducted with free credits provided by AWS educate team.