News
-
Oct 2018:
Shape matching data, training and testing code added in MATLAB/MatConvNet toolbox, available here.
-
May 2018:
Training and evaluation code is now implemented in Python/PyTorch as well, available here.
-
Jul 2017:
Training code is released for MATLAB/MatConvNet framework, available here.
-
Jul 2016:
Data and evaluation code in MATLAB are released.
Papers
-
Deep Shape Matching
Radenović F., Tolias G., Chum O.
ECCV 2018 [ pdf | bib ]
-
Fine-tuning CNN Image Retrieval with No Human Annotation
Radenović F., Tolias G., Chum O.
TPAMI 2018 [ pdf | bib ]
-
CNN Image Retrieval Learns from BoW: Unsupervised Fine-Tuning with Hard Examples
Radenović F., Tolias G., Chum O.
ECCV 2016 [ pdf | bib | poster | presentation | oral ]
2D t-SNE visualization of CNN codes
Downloads
Training Data
Two training sets are provided, comprising 30k and 120k images, with the former being a subset of the latter.
-
Training database: Data used for CNN training with our MATLAB or Python code. It contains the image names lists for training and validation, the cluster ID (3D model ID) for each image and indices forming query-poitive pairs of images. In the case of the 30k dataset the images are all loaded at once and resized in advance to a maximum 362 x 362 dimension, while for the 120k dataset the images are loaded per epoch and resized on the fly to the desired dimensionality. In the case of the edgemap database, the edges are detected in advance on images resized to a maximum 200 x 200 dimension. The original image files are available bellow. Download links:
MATLAB structure:
Python dictionary:
-
Matching pairs for supervised whitening: MATLAB structure used to learn the supervised whitening with our code. It contains query-poitive image pairs. The original image files are available bellow. Download links:
MATLAB structure:
Python dictionary:
-
Images: Original jpeg image files for all 120k images, compressed in tar.gz format, and a MAT-file with splits for the 30k and 120k subsets accompanied by their respective cluster ID. Download links:
Trained Networks
-
Fine-tuned networks with MAC or GeM pooling and contrastive loss:
MATLAB / MatConvNet:
Python / PyTorch:
Code
-
CNN Image Retrieval toolbox implements the training and testing of the approach described in our papers. Toolbox is implemented using MATLAB/MatConvNet and Python/Pytorch frameworks. Data and networks necessary for training and testing are automatically downloaded with example scripts. For detailed description on how to setup and run the code visit: