eseg 1.1.0

Models and utilities for event-based depth / segmentation (Surreal benchmark).

Human Instance Segmentation using Evn

Features

• ConvLSTM-based depth estimation model for event streams
• MobileNetV2 feature encoder with UNet-like decoder
• Event voxelization and augmentation utilities
• Real-time camera viewers (Metavision / DAVIS) with overlay visualization
• Mixed perceptual + edge loss utilities (LPIPS + Sobel)

Requirements

We implemented the dataviewers on Both dvprocessing and metavisionSDK

Please install metavisionSDK for Prophesee live camera.

⚠️ Warning DO NOT forget to set metavisionsdk in your python path especially for windows!

And / Or

dvprocessing for Davis Cameras

For hdf5 file we use metavisionSDK aedat files can be processed using dvprocessing

Pleas, before installing eseg install a GPU enabled pytorch here: https://pytorch.org/get-started/locally/

Installation

pip install eseg

(Once published to PyPI.)

For development:

git clone https://github.com/youruser/eseg.git
cd eseg
python -m venv .venv
source .venv/bin/activate  # Linux / macOS
pip install -e .[dev,viewer]

⚠️ Warning if you work on a virtualenvironment you will need to copy your global sdk library to your local environment

cp -r path/to/your/metavisionsdk/metavision_* <path/to/your/virtualenv/python<yourversion>/site-packages/

Quick Start

import torch
from eseg.models import ConvLSTM
# TODO: usage example after final API stabilizes

Live Stream

python -m eseg.live_stream

Testing

pytest

License

MIT. See LICENSE.

Disclaimer

Research code; APIs may change before 1.0.0.