From d2edf69dbc20577d7bc36d0c72869c11ff0f378c Mon Sep 17 00:00:00 2001 From: Darpan Jain Date: Mon, 18 Feb 2019 18:07:33 +0530 Subject: [PATCH] Update README.md --- README.md | 23 +++++++++++++++++------ 1 file changed, 17 insertions(+), 6 deletions(-) diff --git a/README.md b/README.md index 1b79074..7010b12 100644 --- a/README.md +++ b/README.md @@ -1,5 +1,14 @@ # CROWD COUNTER +## Contents +- [People](#introduction) +- [Nature](#nature) +- [Objects](#objects) +- [Places](#places) +- [Symbols](#symbols) +- [Uncategorized](#uncategorized) + + ### Introduction This repository contains the code of performing the task of implementing a people counter from an overhead video surveillance camera by using Transfer Learning. @@ -13,25 +22,23 @@ This repository contains the code of performing the task of implementing a peopl

-- Clearly, using the pre-trained models is not the way to go. So, ~~Pre-Trained Models~~ +- Clearly, using the pre-trained models is not the way to go. So, ~~Pre-Trained Models~~ :confused: ### Training your custom model A custom model had to trained for accurate implementation. The following steps were taken for the same: -1. **Annotated training data** had to be prepared before being given to the model. +1. **Annotated training data** had to be prepared before being given to the model fot training. 1. ***LabelImg*** was used for this purpose, to draw bounding boxes around objects of interest in the training images. 1. LabelImg gives output as **xml** files containing coordinates of each of the bounding boxes in an image and the associated label of the object. 1. All the xml files were converted to a ***train.csv*** and then into a ***train.record*** format. TFRecord format is required by Tensorflow to perform training of a custom model. 1. Similarly a ***val.record*** was created for validation data. 1. The architecture of the model is based on the **Faster RCNN** algorithm, which is an efficient and popular object detection algorithm which uses deep convolutional networks. -1. The model config file was modified for the purpose of this assignment. -1. The last 90 neuron classification layer of the network was removed and replaced with a new layer that gives output for only one class i.e. person. +1. The config file of the model was modified. The last 90 neuron classification layer of the network was removed and replaced with a new layer that gives output for only one class i.e. person. 1. The **config file** for the same can be found in `./data/utils/faster_rcnn.config` 1. After training the model, the checkpoint model is saved as `model.pb` file. -1. This model can now be deployed and used for obtaining inferences -1. The test images were the first 10 frames of the mall dataset. +1. This model can now be deployed and used for obtaining inferences on crowd images. - The model can be found on this drive link: ​[Custom Model](https://drive.google.com/open?id=1IBgEyaASf10KUFTCbky9mtruUpyoqDWR) @@ -42,9 +49,13 @@ Upon running `main.py`, the results are as shown below. (Refer `./results`)

+ +

+

***Note:*** Since the model was trained on only **30** annotated images, the accuracy can be significantly increased by using a larger dataset to build the model. + ### Prerequisites All the required dependencies can be install by running the command `pip install -r requirements.txt`