We provide the EfficientNet-B4 checkpoints for the reproduced s2(M,f*) and our SemP(100,125,250, f) variants including the respective partitioning and class mapping to the training data:
export dl_path="data/wacv22_checkpoints_efficientnet"
mkdir -p $dl_path/semp_100_125_250
wget https://tib.eu/cloud/s/F5o6MP7y7AbLSpi/download/semp_100_125_250.zip -O $dl_path/semp_100_125_250.zip
unzip $dl_path/s2.zip -d $dl_path/semp_100_125_250/
mkdir -p $dl_path/s2
wget https://tib.eu/cloud/s/4EYyook2nm6EsNQ/download/s2.zip -O $dl_path/s2.zip
unzip $dl_path/s2.zip -d $dl_path/s2/export baseckpt="data/wacv22_checkpoints_efficientnet/s2"
python geo_classification/evaluate_testsets.py --config $baseckpt/hparams.yaml --checkpoint $baseckpt/base.ckpt
export baseckpt="data/wacv22_checkpoints_efficientnet/semp_100_125_250"
python geo_classification/evaluate_testsets.py --config $baseckpt/hparams.yaml --checkpoint $baseckpt/base.ckpt
| testset | checkpoint | acc@1km | acc@25km | acc@200km | acc@750km | acc@2500km |
|---|---|---|---|---|---|---|
| Im2GPS3k | SemP({100, 125, 250}, f) | 12.5 | 31.4 | 42.7 | 57.3 | 72.0 |
| Im2GPS3k | s2(M, f*) | 11.7 | 31.5 | 41.9 | 56.1 | 70.6 |
| Im2GPS | SemP({100, 125, 250}, f) | 16.9 | 45.6 | 57.4 | 72.6 | 85.7 |
| Im2GPS | s2(M, f*) | 15.6 | 41.8 | 54.9 | 71.3 | 81.4 |
| yfcc4k | SemP({100, 125, 250}, f) | 9.4 | 20.3 | 30.5 | 44.8 | 61.3 |
| yfcc4k | s2(M, f*) | 7.5 | 19.7 | 28.5 | 42.5 | 59.1 |
Download and prepare testsets (Im2GPS, Im2GPS3k, YFCC4k):
mkdir -p ../data/images/testsets
cd ../data/images/testsets
mkdir -p im2gps/img
mkdir -p im2gps3k/img
mkdir -p YFCC4k/img
# download and move raw images into their respective folders
wget http://graphics.cs.cmu.edu/projects/im2gps/gps_query_imgs.zip -O im2gps/img2gps.zip
unzip im2gps/im2gps.zip -d .im2gps/img
Im2GPS3k and YFCC4k according to the instructons in https://github.com/lugiavn/revisiting-im2gps/
By default all testsets are evaluated according to the respective hparams.yaml of a model checkpoint using:
python evaluate_testsets.py \
--config ../data/trained_models/<some_configuration>/hparams.yaml \
--checkpoint ../data/trained_models/<some_configuration>/<some_checkpoint>.ckptExpected folder structure: data/images/testsets/<testsetname>/<img>
Expected file format: <testsetname>/*.csv
The header of each CSV file contains at least [img_id, latitude, longitude], where img_id is also the image filename, located in the subdirectory img of the CSV folder.
In general, the training dataloader supplies a batch of images and a list of tensors containing class indices for multi-head classification. Thus, for training each batch has the format, (torch.Size(B, 3, H, W), list([torch.Size(B), torch.Size(B), ...])) while for validation latitude and longitude is concatinated torch.Size(B), torch.Size(B)) to this list to log the great circle distance error also during training.
We use same format as in https://github.com/TIBHannover/GeoEstimation#Training-from-Scratch to store and load the image data and subsequently join additionally required information as class indices or coordinates in MsgPackIterableDatasetMultiTargetWithDynLabels. Please note, that you can write your own dataset class if you don't want to be constrained to our implementation.
See example config files for model and training details.
# supports pytorch lightning trainer flags
python train.py --config configs/example_s2.yml # baseline: s2(M, f*)
python train.py --config configs/example_semp.yml # SemP({100, 125, 250})