Normalisation and Final Score computation for the RL Project

Hi, base of the log operator will be 2.

View post 0

Normalisation and Final Score computation for the RL Project

Each of the kbc assignments will be normalised to 1.

View post 0

Normalisation and Final Score computation for the RL Project

Hi,

The final scores will be computed by adding the scores of each of the 5 environments after normalising these between 0-1.

For the Acrobot and the Taxi environment, min-max normalisation will be used.
For the KBC problems, min-max normalisation after applying a log operator will be used.
The scores will be computed after the submission period has ended.

Thanks,
Siddhartha.

View post 0

Running Evaluations locally

Hi,
In case you are running out of submission limit, please note that the evaluations can be run locally end to end and get an estimate of the scores before making a submission on the leaderboard.
Thanks
Siddhartha.

View post 0

Running local evaluations

Hi,

In case you are running out of submission limit, please note that the evaluations can be run locally end to end and get an estimate of the scores before making a submission on the leaderboard.

Thanks
Siddhartha.

View post 1

New notebook for training with YOLO

We have added a new notebook that walks you through preparing the dataset and configs for training with YOLO models. The notebook also provides an easy interface to apply filters and downloading specific data points for training.

You can find the notebook here:

View post 3

Ideas for getting started

Here is a list of ideas for getting started and improving baselines on the Airborne Object Detection challenge -

Using YOLO models for object detection and tracking

• Out of the box YOLOv3 model with DeepSort algorithm for tracking
• YOLO model fine tuned on the Airprime dataset

Detectron2 provides numerous models for object detection and segmentation and a flexible library for adding new ones. This can be a good starting point for building detection models.

Using JDE based algorithms for object tracking

Most algorithms explored are in SDE (Separate detection and embedding) paradigm. Recent algorithms in the Joint Detection and Embedding (JDE) paradigm have achieved superior performance in MOT leaderboards. Few notable examples -

• FAIR MOT
  • https://arxiv.org/pdf/2004.01888v5.pdf
  • https://github.com/ifzhang/FairMOT
• JDE
  • https://arxiv.org/pdf/1909.12605.pdf
  • https://github.com/Zhongdao/Towards-Realtime-MOT
• GSDT - JDE+GNN based detection, using graphs to connect objects across temporal and spatial dimensions.
  • https://github.com/yongxinw/GSDT
  • https://arxiv.org/pdf/2006.13164v3.pdf

PaddlePaddle provides a framework with a few MultiObject tracking methods (deepsort, fairmot, jde), and a flexible interface for adding new ones.

A few suggestions for improving object detection

• Downscaling the input resolution to the model might make it hard to detect smaller airborne objects. Scaling up your input resolution to the model could help.
• Using a model with high resolution would make your model large and increase inference time. Alternatively, you could tile your images as a preprocessing step and continue using smaller models. (The Power of Tiling for Small Object Detection)
• Removing birds from predictions can help avoid extra false positives, since we are not interested in alerting birds.

View post 6

About the The Neural-MMO Challenge

This forum is for long-form discussions on Neural MMO that are not suited to a single-thread format. Please direct all other discussion and support requests to the Discord server.

View post 0