DeAttNet: Depth Attention Network

Paper: Depth as attention to learn image representations for visual localization, using monocular images https://doi.org/10.1016/j.jvcir.2023.104012

1. Depth image rendering
   1. Using data_prep/gen_depthmaps.py
   2. Zoedepth is available on torch.hub, So you can either directly use it or use via a local repository
   3. Refer to https://github.com/isl-org/ZoeDepth , for any zoedepth loading issues
2. Training
   1. train.py
   2. Configurations defined in config/train.ini
   3. Provide command line arguments accordingly specially dataset_root_dir
   4. Training supports both Cambridge Landmarks and Mapillary Street-level Sequences datasets
   5. Supports netvlad, max, gem descriptors/pooling.
3. Evaluation
   1. xx_retrieval.py perform the retrieval and create the result file
   2. xx_pos_eval.py gives recall@k and avg.pos.err @k results

Significant parts of the code are based on following repositories https://github.com/mapillary/mapillary_sls https://github.com/QVPR/Patch-NetVLAD