This repository contains Jupyter notebooks created for various tasks - from analyzing the dataset to running experiments - to common code implementing PyTorch datasets and models which are used in the notebooks.
Apart from standard libraries such as numpy, pandas and PyTorch, Detectron2 also needs to be installed. Details can be found here. Some additional details may need to be adjusted on a per machine basis, such as paths to directories.
Below are descriptions for each file.
All notebooks can be found in the notebooks/
folder, except for the GRAM
notebook which, due to its size, could not be uploaded to GitHub but can be found here. Some notebooks will require the download of the dataset here.
- Detectron: contains preliminary experimentation with Detectron2
- GRAM: analysis of the M-30 video sequence from the GRAM-RTM, used for the conversion of the provided bounding box annotations in XML to labels for each frame (count of cars) and manual inspection of correctness of the labels. Warning: around 500MB in size, may take a while to load.
- Data Analysis: analysis of distribution of labels among train, validation and test sets.
- RCNN: preliminary tests on adapting RCNN models to count objects
- Maps: used to generate and store feature maps from backbones to be used with models under
rcnn_models/
for the comparison notebooks - Specialized NNs: preliminary tests on using the specialized NN from BlazeIt. Note: this notebook was created to be used with Google Colab.
- Comparisons: contains tests for performance (subsection 4.1 in the PDF) and usage of different backbones (subsection 4.2 in PDF)
- Comparisons2: contains tests for performance comparison while using Adam optimizer for specialized NN, and correlations (subsection 4.3 in PDF)
- Comparisons3: contains tests for comparing performance with RCNN models (subsection 4.4 in PDF)
gram.py
: PyTorch dataset for M-30 video sequence from GRAM-RTM and labels of counts of carsact_maps.py
: PyTorch dataset used to load activation maps generated inMaps
notebook instead of M-30 framesrcnn.py
: RCNN models from Detectron2 adapted so that the forward pass outputs the count of cars in an imagespecialized_nn.py
: specialized NN code adapted from BlazeItrcnn_models/linear_1_conv_1
: Architecture 1 in PDFrcnn_models/linear_2_conv_1
: Architecture 2 in PDFrcnn_models/linear_1_conv_2
: Architecture 3 in PDFrcnn_models/linear_2_conv_2
: Architecture 4 in PDFrcnn_models/linear_3_conv_3
: Architecture 5 in PDF