Learning all about the amazing life of Saint Francis of Asissi who inspires us all to take care of our common home!

Based on the zoom insights, we take all existing labeled images that OpenAI's CLIP ViT-L/14 never correct labels even given 324 zoom attempts and form a new benchmark called 🌟 ImageNet-Hard 🌟

7/n

Night, night, sleep tight 😴😴😴

For example, zoom-in helps models correctly classify in (junco) and (magpie).
Zoom-out also helps, but mostly on abstract images (e.g. ImageNet-Sketch) (hornbill).

3/n

数日後5、6年振りにででに🐭🏰に行くのでとってもワクワクしてます、、、

Visiting King Charles!

実は私も昔に見つけて😡😡ってなってました〜！！
買う人居ないだろうけど見つけても買わないようにね、、、🥲🥲

Zoom definition: first, uniformly resize, and then crop.

Surprisingly, even 11-year-old AlexNet is good enough to classify > 90% of ImageNet given an *ideal* zoomed crop.

This upper-bound accuracy increases as models are more recent (AlexNet < VGG < ResNet < ViT).

2/n

とりあえず、ファミマのぼっちコラボのファイルを貰ってきた。

#ぼっち・ざ・ろっく

🐦️🐤🐦️

#필릭스 #Felix #StrayKids
#StrayKids Comeback

peace 🐤

#필릭스 #Felix #StrayKids

When you travel, you want to make your time worthwhile. Well, we are here to help steer you away from some bad decisions.

Integrating a zoom-in ⨁ process into MEMO (Zhang, Chelsea Finn Sergey Levine 2022), a state-of-the-art test-time augmentation method, can increase prediction accuracy even further.

6/n

抵抗してみる？😆

By center-zooming to each of the 9 anchor points on every image, we show that ImageNet-A and ObjectNet have a strong center bias ➡ model accuracy increases dramatically (e.g. +15% on IN-A) by simply center-cropping input images.

5/n

牛乳を注ぐプルャ

#ちいかわファンアート

There are ~200 images (samples below) in ImageNet that are *never* correctly classified by any models even if each model is given 324 zoom attempts ❗️

Looking at this *unclassifiable* set of images informs us about 'Are we done with ImageNet-scale OOD benchmarks?'

4/n

How to score > 90% on ImageNet?

Our new study on the spatial biases of ImageNet and relevant ImageNet-scale, OOD benchmarks reveals that all common image classifiers tested can score > 90%, if the model looks at the correct crop, i.e.,

⭐️ Zoom 🔎 is all you need! ⭐️ 1/n

This #healthsystem increases #resilience and #efficiency with #AI and contextualized #data . How Johns Hopkins Medicine addresses #supplychain disruptions head on. Learn more Deloitte. #ad bit.ly/3G1Wv7f

ImageNet-Hard is a new benchmark that challenges state-of-the-art vision-language models and ImageNet-trained classifiers.

Work led by the amazingly talented taesiri!!! With Giang Nguyen , Sarra Habchi, and Cor-Paul Bezemer.

Code, paper & dataset: taesiri.github.io/ZoomIsAllYouNe…

n/n

👍 You’ll Never Guess Why People Loves These Backpack so Much!
⭐ Airline Carry-On Size ⭐
✅ 15.6 Laptop Compartment
✅ Big Capacity
✅ Combining Style With Functionality

On ImageNet-Hard, all 224x224-input models perform below 19%.

Higher-resolution image classifiers tend to perform better with the best SOTA of 39% by EfficientNet-L2@800px.
Manually increasing the resolution of input images by GigaGAN does not improve model accuracy.

8/n

Year 2 enjoyed a fascinating visit to the Gurdwara this week.