MLDL-Final-Project

This project propose some extensions to improve Deep Visual Geo-Localization Benchmark ued for the task of VPR. The proposed extensions improve robustness to night images and occlusions, apply new optmizers for training the model, propose a multi-scale strategy for building richer descriptors and provide a way to combine models. Results are in Results Analysis.xlsx Excel file.

Train Improvements

Data Augmentation Night Robustness

• datasets_augmented.py: in this file we have define the data augmentation for night robustness with ColorJitter.
• train_augmented.py: we implement the train loop with this Color Jitter transformation.
• datasets_constant.py: in this file we have define the data augmentation for night robustness with Functional Adjust.
• train_augmented_constant.py: we implement the train loop with this Functional Adjust transformation

Data Augmentation Perspective

• datasets_RP_RE.py: in this file we have defined data augmentation for occlusions with Random Perspective and Random Erasing.
• train_RP_RE.py: we implement the train loop with these Random Perspective and Random Erasing transformations.

Optimizers and Scheduler

• train_optim.py: in this file we have defined the training loop with the new optimizers: ADAMW and ASGD
• train_optim_sched.py: in this file we have defined the training loop with two schedulers: Cosine and Plateau.

Test Improvements

Multi-scale testing

• test_impr.py: in this file we have define the multi-scale pyramid

Ensemebling

• test_ensam.py: in this file we have defined the code for ensembling descriptors from different models
• eval_impr.py: in this file we implement both multi-scale and ensembling extensions for running it in the command line

Vanilla model

• commons.py, datasets_ws.py, eval.py, parser.py, test.py, train.py and util.py and the folder model are the files from the original model.

Project Description

Data Augmentation.png

Multi-scale.png

Ensembling.png