The colloquium at CTU in Prague had 6 great talks and a lot of discussions before, during and after the event. The slides are now shared online. It was the 50th and our administrators surprised us with a huge Czech cake - Koláč. See you in April again!

The Visual Recognition Group at CTU in Prague organizes the 50th Pattern Recognition and Computer Vision Colloquium with
Torsten Sattler, Paul-Edouard Sarlin, Vicky Kalogeiton, Spyros Gidaris, Anna Kukleva, and Lukas Neumann.
On Thursday Oct 9, 11:00-17:00.

cmp.felk.cvut.cz/colloquium/

Better than just hope: in our ECCV2020 we show that GAP (global average pooling) is a good way to optimize local descriptors using image-level supervision and loss. The dino.txt paper switches from the CLS token to GAP and look at the segmentation task improvements.

A CVL is better able to predict the image processing or acquisition settings given solely the image representation when it is trained without image augmentations.

growing audience

The Visual Recognition Group at CTU in Prague organizes the 49th Pattern Recognition and Computer Vision Colloquium with D. Karatzas, M. Masana, T. Tommasi, P. Mettes @pascalmettes.bsky.social , E. Brachmann @ericbrachmann.bsky.social and V. Stojnic @stojnicv.xyz

cmp.felk.cvut.cz/colloquium/#...

Clutter is a major challenge. Small objects are ranked very low using global representations. Oracle experiment: we mask out the background or crop-resize the object and observe a big space for improvements if you can handle heavy clutter.

It is striking that off-the-shelf use of VLMs shows that particular objects can be retrieved with textual queries nearly as effectively as with image queries

CVPR

history

from a paper - not a joke

in a comparison between using frozen DINOv2 and a model trained for the specific domain in a supervised way (SG/CVNet): DINOv2 is much worse as a global descriptor, but much better for local descriptors.

AMES is the new state-of-the-art in instance-level retrieval, has a better performance-memory trade-off than existing models and can improve global descriptor performance even with as few as 10 128D binary local descriptors

Transformers are sensitive to sequence length discrepancy between training and testing. Training AMES for a fixed number of local descriptors and testing it for a different one shows performance losses. We eventually train a universal model with a varying number of descriptors

Architecture: image2image similarity via a binary classifier that processes two vector sets with within image and across image attention. Memory savings: more local descriptors for the query and less for the database image + use binary descriptors at a tiny performance drop

Deep global descriptors give a convenient way for retrieval, but local descriptors are a game changer in finding needles in a haystack (particular objects in clutter). Due to their high cost, with AMES we optimize the performance/memory trade-off during re-ranking. #ECCV2024