evalYFCC

Evaluation on Two-View Geometry Estimation

Data

We follow the evaluation of OANet. Downloading the 4 test scenes used in OANet : buckingham_palace , notre_dame_front_facade, reichstag, sacre_coeur.

Once it is downloaded, saved it into ../../data, the file structure should be :

./RANSAC-Flow/data/YFCC

├── images/
│   ├── buckingham_palace/
│   ├── notre_dame_front_facade/
│   ├── reichstag/
│   └── sacre_coeur/
└── pairs/
    ├── buckingham_palace-te-1000-pairs.pkl
    ├── notre_dame_front_facade-te-1000-pairs.pkl
    ├── reichstag-te-1000-pairs.pkl
    └── sacre_coeur-te-1000-pairs.pkl

MOCO feature

SegNet + W/O Fine-tuning

Running :

python evaluation.py --outDir MOCO --segNet YFCC 

To get results with our fine alignment :

python getResults.py --multiH --ransac --coarsePth MOCO_Coarse --finePth MOCO_Fine --maskPth MOCO_Fine --outRes moco.json --scene 0
python getResults.py --multiH --ransac --coarsePth MOCO_Coarse --finePth MOCO_Fine --maskPth MOCO_Fine --outRes moco.json --scene 1
python getResults.py --multiH --ransac --coarsePth MOCO_Coarse --finePth MOCO_Fine --maskPth MOCO_Fine --outRes moco.json --scene 2
python getResults.py --multiH --ransac --coarsePth MOCO_Coarse --finePth MOCO_Fine --maskPth MOCO_Fine --outRes moco.json --scene 3

ImageNet feature

Adding --imageNet when running evaluation.py with the above commands.

Metric

According to the implementation of OANet, the mAP@20 in the paper is the average over AP < 5, AP < 10, AP < 15, AP < 20; mAP@10 in the paper in the average over standard Acc < 5, Acc < 10. For more details, we refer to this part of code in OANet.

Running our code will give you the AP < 5, AP < 10, AP < 15, AP < 20 (average over the 4 scenes). Then to compare to the numbers in the paper, one need to compute the average.

We thank to Jiahui Zhang to point it out.