This repository is based on CosPlace project available in the link below:

[https://github.com/gmberton/CosPlace.git].

In this repository all the training and testing are done using sf_xs and tokyo_xs.

1. sf_xs size is 1,22 GB on disk and 103.869 items.

2. tokyo_xs size is 164,6 MB on disk and 13.090 items.

In addition two backbones and data augmentation is included:

backbones	augmentation
Wide ResNet50-2	use_horizontal_flip
densenet121	degrees

---

Results of our tests on the datasets are shown as below:

Best models results: Recall@1 / Recall@5

SF-XS val(R@1 / R@5)	SF-XS(R@1 / R@5)	Tokyo-XS(R@1 / R@5)
56.4 / 69.4	20.9 / 32.6	34.0 / 52.4

---

Results of different parameter for data augmentation:

Model	SF-XS val(R@1 / R@5)	SF-XS(R@1 / R@5)	Tokyo-XS(R@1 / R@5)
degree=5	56.6 / 69.5	22.2 / 35.0	33.3 / 54.3
degree=10	56.0 / 68.7	22.3 / 36.7	30.8 / 56.2
horizontal_flip=true	55.9 / 69.6	19.7 / 32.6	31.1 / 54.3

---

Comparison between ResNet18 and VGG16: Recall@1 / Recall@5

Model	SF-XS val(R@1 / R@5)	SF-XS(R@1 / R@5)	Tokyo-XS(R@1 / R@5)
ResNet18	56.4 / 69.4	20.9 / 32.6	34.0 / 52.4
VGG16	70.0 / 80.8	36.5 / 50.7	51.7 / 74.0

---

Comparison of Performance between Adam and AdamW Optimizers

Dataset	Default (Adam lr e-5)(R@1 / R@5)	AdamW (lr e-5)(R@1 / R@5)
SF-XS val	56.4 / 69.4	56.2 / 69.3
SF-XS test	20.9 / 32.6	19.6 / 32.5
Tokyo-XS	34.0 / 52.4	33.3 / 53.3

---

---

Our tests and pre-trained models are available at this link:

[https://drive.google.com/drive/folders/1--7FrsGjZBLAnE02On0ekd9endqXB3lE?usp=sharing]