F2DNet reproduction

This paper was reproduced as part of the project in CS4245 course in TU Delft CS Masters.

You can find the original GitHub repository here.

We aim to reproduce the results independently, without use of the original code. But because it's a really large project and cite lots of existing packages, we are not able to reproduce it from scratch fully, but we did write some parts from sractch and produce some results. Also, we can run their official code to produce some outputs.

1. DATASET - CityPersons dataset loading

1. Download the CityPersons dataset
   • The structure should be as shown below

├── datasets
│   ├── CityPersons
│     └── annotations
│         ├── anno_train.mat
│         └── anno_val.mat
│     └── images
│         ├── train
│         ├── val
│         └── test

Note

DataLoading code taken from lwpyr/CSP-pedestrian-in-pytorch. We plan to change that as well.

2. BACKBONE - HRNet reproduction(from scratch)

Dependencies:

1. pytorch(the stable vision in official website is fine)
2. The pretrained model should be downloaded from https://drive.google.com/file/d/1NxCK7Zgn5PmeS7W1jYLt5J9E0RRZ2oyF/view and put in the HRNet_pretrain/checkpoints folder
3. The original paper can be found in https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9052469

Get started

To verify the results, go to the HRNet_pretrain folder, run the hrnetpretrain.py file, the input size is 2048*1024, and the output tensor size is [1, 480, 256, 128].

The HRNet is the backbone of our F2DNet. Its goal is to represent the network input in a semantically richer and spatially precise way.

The detailed documentation for HRNet can be found in HRNet_pretrain.

3. DETECTION -