visual_localization

Visual localization code suitable for Hyundai dataset in naver labs challenge.

visual_localization contains 3 parts:

• extract_features.py
• eval.py
• eval_superglue.py

extract_features.py

This part extracts global descriptor of a image.

usage: extract_feature.py [-h] [--floor FLOOR] [--globaldesc GLOBALDESC] [--batchsize BATCHSIZE] 
                          [--dataset_path DATASET_PATH] [--checkpoint_path CHECKPOINT_PATH]

Arguments:

• --floor: Which floor: 1f or b1?
• --globaldesc: Global descriptor to use: netvlad or apgem
• --batchsize: Batch size
• --dataset_path: Path to dataset directory (parent directory of 1f and b1)
• --checkpoint_path: Path to checkpoint

Example:

python extract_feature.py --floor b1 --globaldesc netvlad --batchsize 32 \
--dataset_path ../dataset --checkpoint_path ../checkpoints/vgg16_netvlad_checkpoint/checkpoints/checkpoint.pth.tar

It saves global descriptor of each db and query image as 2d numpy array. The saved file name is {globaldesc}_db_{floor}_features.npy and {globaldesc}_query_{floor}_features.npy

eval.py

This part performs pose estimation using rootSIFT. It need result of extract_features.py.

usage: eval.py [-h] [--floor FLOOR] [--globaldesc GLOBALDESC] [--rank_knn_num RANK_KNN_NUM] 
               [--rerank_knn_num RERANK_KNN_NUM] [--dataset_path DATASET_PATH]

Arguments:

• --floor: Which floor: 1f or b1?
• --globaldesc: Global descriptor to use: netvlad or apgem
• --dataset_path: Path to dataset directory (parent directory of 1f and b1)
• --rank_knn_num: Number of nearest neighbor to select at ranking phase
• --rerank_knn_num: Number of nearest neighbor to select at reranking phase

Example:

python eval.py --floor b1 --globaldesc apgem --rank_knn_num 10 \
--rerank_knn_num 5 --dataset_path ../dataset

It saves answer as json file with name {globaldesc}_rootsift_rank_knn_num_{rank_knn_num}_rerank_knn_num_{rerank_knn_num}_answer_{floor}.json. It also saves intermediate result on every 100 image evaluations.

eval_superglue.py

This part performs pose estimation using SuperGlue. It need result of extract_features.py.

usage: eval_superglue.py [-h] [--floor FLOOR] [--globaldesc GLOBALDESC] [--rank_knn_num RANK_KNN_NUM] 
               [--rerank_knn_num RERANK_KNN_NUM] [--dataset_path DATASET_PATH]

Arguments:

• --floor: Which floor: 1f or b1?
• --globaldesc: Global descriptor to use: netvlad or apgem
• --dataset_path: Path to dataset directory (parent directory of 1f and b1)
• --rank_knn_num: Number of nearest neighbor to select at ranking phase
• --rerank_knn_num: Number of nearest neighbor to select at reranking phase

Example:

python eval_superglue.py --floor b1 --globaldesc apgem --rank_knn_num 10 \
--rerank_knn_num 5 --dataset_path ../dataset

It saves answer as json file with name {globaldesc}_superglue_rank_knn_num_{rank_knn_num}_rerank_knn_num_{rerank_knn_num}_answer_{floor}.json. It also saves intermediate result on every 100 image evaluations.