🧐7 things you should know about the Focal Loss:

📌 It was introduced in the RetinaNet paper to address the foreground-background class imbalance encountered during training of dense detectors (one-stage detectors)
...
📌 It’s derived from the cross-entropy loss such that it down-weights the loss assigned to well-classified examples. It's used in the classification head.

📌 It’s used in many one-stage object detection models: EfficientDet, FCOS, VFNet, and many other models
📌 It can also be used in two-stage object detection models: e.g. Sparse R-CNN

📌 It crashes losses associated to easy examples: for a confidence score of 0.9, the focal loss is 100 times smaller than the cross-entropy loss (see figure here above)
📌 Thanks to the focal loss, RetinaNet was the first one-stage detector model to beat two-stage detector models

📌 Focal loss can also be used in classification centered tasks (not only in object detection tasks)

🤔 Are there other things you would add to this list?
Thanks for passing by!

🟧Def Follow @ai_fast_track for more stuff on Object Detection

🟦and if you could give the thread a quick retweet, it will help other people catch this content in their feed 🙏

• • •

Missing some Tweet in this thread? You can try to force a refresh
 

Keep Current with AI Fast Track

AI Fast Track Profile picture

Stay in touch and get notified when new unrolls are available from this author!

Read all threads

This Thread may be Removed Anytime!

PDF

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

Try unrolling a thread yourself!

how to unroll video
  1. Follow @ThreadReaderApp to mention us!

  2. From a Twitter thread mention us with a keyword "unroll"
@threadreaderapp unroll

Practice here first or read more on our help page!

More from @ai_fast_track

27 Oct
✨Common Object Detector Architecture you should be familiar with:

📌 Common object detectors are divided into One-Stage Detectors (OSD), and Two-Stage Detectors (TSD)

📌 Both OSD and TSD can be either anchor-based (relying on anchor boxes) or anchor-free Image
📌 OSD use the whole feature maps to predict bounding boxes/labels: Dense Prediction

📌 TSD have an extra step hence two-stage: extracting proposals (regions of interest)

📌 Proposals are used to extract feature map regions to predict bounding boxes/labels: Sparse Prediction
📌 TSD don't use the whole feature map for prediction

📌 TSD (e.g. Faster R-CNN) used to be more accurate than STD (e.g. SSD, YOLO, etc.)

📌 STD (e.g. EfficientDet, RetinaNet, VFNet, YOLOX, etc.) recently show better results than TSD

📌 STD are faster than TSD
Read 5 tweets
21 Oct
🥇YOLOX beat YOLOv5. They won the 1st Place on Streaming Perception Challenge (Workshop on Autonomous Driving at CVPR 2021).

I briefly explain how they created it, here below👇

YOLOX will be soon supported in our IceVision Framework (Repo: github.com/airctic/IceVis…)
Considering YOLOv4 and YOLOv5 might be a little bit over-optimized for the anchor-based pipeline, they choose YOLOv3 as their starting point.

They transformed YOLOv3 into YOLOX by using the following techniques:

🟦 Decoupled head,
🟦 Anchor-free architecture (like FCOS), and
🟦 Advanced label assigning strategy (SimOTA)

They also used some other training strategies such as: adding EMA weights updating, cosine LR schedule, IoU loss, and IoU-aware branch.

pdf: arxiv.org/pdf/2107.08430…
abs: arxiv.org/abs/2107.08430
repo: github.com/Megvii-BaseDet…
Read 4 tweets
20 Oct
Here is a cool approach for effective labeling from @NVIDIAAI

📰Active Learning for Deep Object Detection via Probabilistic Modeling

Object detection labeling is both time consuming and expensive. Active learning to the rescue:
🧵 Image
🟧They use mixture density networks to estimate, in one forward pass of a single model, both localization and classification uncertainties

🟧 They leverage them in the scoring function for active learning.

🟧 Then, those data points with the top-K scores are sent for labeling
🟧 They use SSD pytorch for their implementation

📰Paper: Active Learning for Deep Object Detection via Probabilistic Modeling

PDF: arxiv.org/abs/2103.16130
Abs: arxiv.org/pdf/2103.16130…
Repo: github.com/NVlabs/AL-MDN
Read 4 tweets
30 Dec 20
CornerNet Follow-up Paper- CenterNet: Keypoint Triplets for Object Detection
🧵
- CornerNet focus on detecting object edges leads to generating boxes sharing similar edges👇

- 🔥CenterNet Solution🔥: It adds a Center Pooling
- Objects central parts have richer features (map) than corner regions (those rely on Corner Pooling to compensate features lack). CenterNet correctly detects objects by checking the central parts, in addition to their corners.
Performance Comparison: CenterNet ranks among the top SOTA Two-Stage Detectors.
Read 6 tweets
17 Dec 20
Interested in Segmentation with 🔥only🔥 Bounding-Box annotations?

We are starting a Discussion Group in our Discord Forum.

I will cover the 3 papers (Tian, et al.) that the BoxInst architecture is built on, in several presentations (thread👇)

Forum: discord.gg/JDBeZYK
📅Dates will be announced soon. Stay tuned!

- FCOS: Fully Convolutional One Stage Detector (Apr 2019):

Bounding boxes detector without using anchor boxes nor Region Proposal Network (RPN)

arxiv.org/pdf/1904.01355
- CondInst: Conditional Convolutions for Instance Segmentation (March 2020):

It's a per-instance segmentation. They generate specific light filters per instance as opposed to filters common to all instances

arxiv.org/pdf/2003.05664
Read 4 tweets

Did Thread Reader help you today?

Support us! We are indie developers!


This site is made by just two indie developers on a laptop doing marketing, support and development! Read more about the story.

Become a Premium Member ($3/month or $30/year) and get exclusive features!

Become Premium

Too expensive? Make a small donation by buying us coffee ($5) or help with server cost ($10)

Donate via Paypal Become our Patreon

Thank you for your support!

Follow Us on Twitter!

:(