leftthomas / DaCo

A PyTorch implementation of DaCo based on APIN 2023 paper "DaCo: Domain-Agnostic Contrastive Learning for Visual Place Recognition"

DaCo

A PyTorch implementation of DaCo based on APIN 2023 paper DaCo: Domain-Agnostic Contrastive Learning for Visual Place Recognition.

Requirements

• Anaconda
• PyTorch

conda install pytorch=1.7.0 torchvision cudatoolkit=11.0 -c pytorch

Dataset

Tokyo 24/7, Cityscapes FoggyDBF and Synthia Seqs datasets are used in this repo, you could download these datasets from official websites, or download them from MEGA. The data should be rearranged, please refer the paper to acquire the details of train/val split. The data directory structure is shown as follows:

├──tokyo
  ├── original (orignal images)
      ├── domain_a (day images)
          ├── train
              ├── day_00001.jpg
              └── ...
          ├── val
              ├── day_00301.jpg
              └── ...
      ├── domain_b (night images)
          ├── train
              ├── night_00001.jpg
              └── ...
          ├── val
              ├── night_00301.jpg
              └── ...
  ├── generated (generated images)
      same structure as original
      ...
├──cityscapes
  same structure as tokyo
  ...
├──synthia
  same structure as tokyo 
  ... 

Usage

python main.py --data_name synthia --method_name simclr
optional arguments:
--data_root                   Datasets root path [default value is 'data']
--data_name                   Dataset name [default value is 'tokyo'](choices=['tokyo', 'cityscapes', 'synthia'])
--method_name                 Method name [default value is 'daco'](choices=['daco', simsiam', 'simclr', 'moco', 'npid'])
--hidden_dim                  Hidden feature dim for projection head [default value is 512]
--temperature                 Temperature used in softmax [default value is 0.1]
--batch_size                  Number of images in each mini-batch [default value is 16]
--iters                       Number of bp over the model to train [default value is 10000]
--ranks                       Selected recall [default value is [1, 2, 4, 8]]
--save_root                   Result saved root path [default value is 'result']
--lamda                       Lambda used for the weight of soft constrain [default value is 0.8]
--negs                        Negative sample number [default value is 4096]
--momentum                    Momentum used for the update of memory bank or shadow model [default value is 0.5]

For example, to train moco on cityscapes:

python main.py --data_name cityscapes --method_name moco --batch_size 32 --momentum 0.999

to train npid on tokyo:

python main.py --data_name tokyo --method_name npid --batch_size 64 --momentum 0.5

Benchmarks

The models are trained on one NVIDIA GTX TITAN (12G) GPU. Adam is used to optimize the model, lr is 1e-3 and weight decay is 1e-6. batch size is 16 for daco, 32 for simsiam, simclr and moco, 64 for npid. momentum is 0.999 for moco, 0.5 for npid, other hyper-parameters are the default values.

Tokyo 24/7

Method	Day --> Night				Night --> Day				Day <--> Night				Download
	R@1	R@2	R@4	R@8	R@1	R@2	R@4	R@8	R@1	R@2	R@4	R@8
NPID	6.67	10.67	17.33	28.00	8.00	12.00	16.00	21.33	2.00	4.00	5.33	10.67	r2bg
MoCo	5.33	6.67	12.00	17.33	6.67	9.33	12.00	16.00	0.00	0.00	0.00	0.67	f2jt
SimCLR	25.33	32.00	45.33	56.00	33.33	37.33	46.67	58.67	8.67	9.33	14.00	18.67	agdw
SimSiam	4.00	5.33	9.33	16.00	4.00	5.33	6.67	14.67	1.33	1.33	1.33	3.33	d2i4
DaCo	61.33	68.00	78.67	84.00	60.00	70.67	81.33	88.00	45.33	56.67	64.00	74.67	5dzs

Foggy Cityscapes

Method	Clear --> Foggy				Foggy --> Clear				Clear <--> Foggy				Download
	R@1	R@2	R@4	R@8	R@1	R@2	R@4	R@8	R@1	R@2	R@4	R@8
NPID	5.60	7.80	10.20	17.20	5.20	8.40	12.60	19.80	0.20	0.50	0.70	1.00	bbiv
MoCo	0.40	0.80	1.40	3.00	0.40	0.80	1.80	3.40	0.20	0.20	0.20	0.20	ma2a
SimCLR	34.60	48.00	59.60	71.40	54.60	66.80	78.40	86.00	0.50	1.00	1.70	2.90	hdhn
SimSiam	0.60	1.00	1.60	2.40	0.80	1.00	1.80	2.80	0.00	0.00	0.00	0.00	dau5
DaCo	98.80	99.60	99.80	100.0	98.60	99.20	99.40	99.80	90.20	96.00	98.60	99.20	azvx

Synthia Seqs

Method	Sunset --> Rainy Night				Rainy Night --> Sunset				Sunset <--> Rainy Night				Download
	R@1	R@2	R@4	R@8	R@1	R@2	R@4	R@8	R@1	R@2	R@4	R@8
NPID	8.33	10.00	11.67	20.00	6.67	6.67	18.33	21.67	5.00	5.00	10.00	11.67	bgua
MoCo	5.00	5.00	8.33	16.67	10.00	15.00	16.67	25.00	1.67	1.67	1.67	3.33	sw7f
SimCLR	23.33	30.00	46.67	51.67	25.00	33.33	48.33	63.33	6.67	10.83	16.67	20.83	afeg
SimSiam	10.00	13.33	15.00	23.33	5.00	10.00	21.67	30.00	3.33	5.00	6.67	8.33	bjnj
DaCo	55.00	68.33	75.00	88.33	41.67	51.67	70.00	83.33	25.00	34.17	50.00	66.67	sasq

Results

Tokyo 24/7

Foggy Cityscapes

Synthia Seqs

T-SNE