Tweet

@kaggle

@ai_fast_track

This Thread may be Removed Anytime!

Twitter may remove this content at anytime! Save it as PDF for later use!

More from @ai_fast_track

AI Fast Track (36/45)

@ai_fast_track

6 Dec

@ai_fast_track

🎉Part 2- Summary of 10 summaries on:

Tips & Trick & Best Practices in training (not only) object detection models.

Don't miss any of those posts, follow @ai_fast_track to catch them in your feed.

🎁 Summary of summaries: ...

https://twitter.com/ai_fast_track/status/1458493806362378244

1- Training Object Detection Models Tips & Tricks

https://twitter.com/ai_fast_track/status/1458493806362378244

https://twitter.com/ai_fast_track/status/1466998473312677894

2- Pro Tip to fast track your object detection training

https://twitter.com/ai_fast_track/status/1466998473312677894

Read 12 tweets

AI Fast Track (36/45)

@ai_fast_track

4 Dec

@fastdotai

💡 Prop Tip to 🚀 fast track your object detection training

All @fastdotai users use this method for free!

It's fine_tune()

We noticed an important increase in the COCO Metric when we train object detection models using this method even with the simple default values

Fine tuning has 2 phases:

🧊 Phase 1: Freeze your model's backbone and neck. Train your model for a couple of epochs (e.g. 1 to 5)

📈 Phase 2: Unfreeze the whole model. Train for higher number of periods (depending on your dataset, and how hard/easy your data is)

Phase 1 is a kind of warm-up for your model head.

Here is an example showing how to call the fine_tune() method.

In this case, we freeze both the backbone and the neck for 1 epoch in phase 1, and train for 20 epochs in phase 2.

Read 5 tweets

AI Fast Track (36/45)

@ai_fast_track

2 Dec

Detecting small object detection is one of the most challenging tasks in object detection.

❓ Want to know an effective way to overcome that❓

Check out the SAHI integration in IceVision🧊
👇

All IceVision models transparently support SAHI (MMDetection, EfficientDet, Torchvision, and YOLOv5). The results are quite impressive!

SAHI is a very cool library that drastically improves detection of small objects.

📌 It is simply used at inference time

📌 It leverages high resolution images.

📌 Your training process remains the same as usual.

🔎 At the high level, this is how SAHI works:

1- It takes a high resolution (e.g.1920 x 1024 px), and divide it in patches (e.g. 128 x 128 px)

Read 7 tweets

AI Fast Track (36/45)

@ai_fast_track

30 Nov

🎉🎊🥳Day 30/30: Labeling data and training object detection models are time consuming and expensive.

Here is a Survey of Self-Supervised and Few-Shot Object Detection (FSOD). Those technique aim at alleviating those issues.

The authors have categorized, reviewed, and compared several few-shot and self-supervised object detection methods

📌 FSOD is about training a model on novel (unseen) object classes with little data. It still requires prior training on many labeled data of base (seen) classes

📌 Self-Supervised Learning (SSL) methods aim at learning representations from unlabeled data which transfer well to downstream tasks such as object detection

📌 Combining few-shot and self-supervised object detection is a promising research direction

Read 8 tweets

AI Fast Track (36/45)

@ai_fast_track

26 Nov

Day 26/30: General Gaussian Heatmap Labeling is Arbitrary-Oriented Object Detection AOOD

📌 It uses an anchor-free object-adaptation label assignment strategy to deﬁne positive candidates based on 2D GH, reﬂecting shape and direction features of arbitrary-oriented objects.

📌 GGHL improves the AOOD performance with low parameter-tuning and time costs.

📌 It is also applicable to most AOOD methods to improve their performance including lightweight models.

📰 Paper: A General Gaussian Heatmap Labeling for Arbitrary-Oriented Object Detection

abs: arxiv.org/abs/2109.12848…
pdf: arxiv.org/pdf/2109.12848…

Read 4 tweets

AI Fast Track (36/45)

@ai_fast_track

24 Nov

Day 24/30: 🥇 EfficientDet is a very popular object detection model for a good reason!

Let’s see why

📌 EfficientDet achieved State-Of-The-Art (SOTA) accuracy while reducing both the size of parameters, and the FLOPS, when it was released. It’s still a very good contender.

📌 Before introducing EfficientDet, models were getting impressively big to achieve SOTA results

❓ The authors asked the following question:
Is it possible to build a scalable detection architecture with both higher accuracy and better efficiency across # resource constraints?

So, they systematically studied neural network architecture design choices for object detection, and proposed several key optimizations to improve efficiency:

1- A weighted bi-directional feature pyramid network (BiFPN), which allows easy and fast multiscale feature fusion

Read 8 tweets

Share this page!

AI Fast Track (36/45)

Try unrolling a thread yourself!

More from @ai_fast_track

AI Fast Track (36/45)

AI Fast Track (36/45)

AI Fast Track (36/45)

AI Fast Track (36/45)

AI Fast Track (36/45)

AI Fast Track (36/45)

Did Thread Reader help you today?

Like this author's thread?