Hacker News new | past | comments | ask | show | jobs | submit login
PIRL: Learn Image Representations Immune to Geometric Transformation (amitness.com)
50 points by amitness on March 17, 2020 | hide | past | favorite | 6 comments



Thank you @amitness for this wonderful website with graphical explanations :)

The PIRL technique in question here seems useless to me, because its loss deliberately trains it to behave differently from what a human would do. But that overview page https://amitness.com/2020/02/illustrated-self-supervised-lea... is gold.


This looks a lot to me like learned image descriptors from the computer vision community. Some examples are Discriminative learning of local image descriptors (http://matthewalunbrown.com/papers/pami2010.pdf) DeepDesc (https://icwww.epfl.ch/~trulls/pdf/iccv-2015-deepdesc.pdf) L2-net (http://www.nlpr.ia.ac.cn/fanbin/pub/L2-Net_CVPR17.pdf)


The issue with this is that you (most of the time) don't want your image representations to be immune to geometric transformation.

A rotated p should be recognized as a d. In nature pictures you want to recognize blue at the top as sky and blue at the bottom as water.


I think that's taking a dim view of what machine vision should be capable of. If a picture is rotated 180 degrees, upside down text should be recognized as its flipped version, i.e., the word "6op" should be read as "dog" having been rotated. Similarly a machine should be able to tell if it's looking at the reflection of the sky in a lake. And disregarding entire scenes, any object upside down (upside down chair, upside down dog, upside down car, ...) is clearly recognizable as such to a human, and hence should be to a machine as well.


Upside down objects and scenes look alien to us. We can recognize them but not without effort. It's a combination of having seen objects upside down before and being able to reason about what we are seeing. It does not come for free from invariance.


The classic Thatcher illusion[0].

[0] https://en.wikipedia.org/wiki/Thatcher_effect

Edited for clarity, started writing then wiped half and didn't check first half.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: