nishaq503/rheapose

RheaPose

Benchmarks

This repository uses git submodules to include the Vector-Field converter tools from PolusAI. You will need to update the submodules to get the latest version of the tools:

git submodule update --init

We used Python 3.11.8 on a Linux to run all the benchmarks. The benchmarks were run using a single CPU core and no GPU.

To get started, create a virtual environment and install the package with poetry:

python -m venv .venv
source .venv/bin/activate
poetry install

Then, you can run the benchmarks with:

python benches/bench_vector_converters.py

Note that the Omnipose package is currently broken and will not run.

TissueNet Data

1. Download the TissueNet dataset (v1.1) from DeepCell Datasets.
2. Extract the dataset to a location of your choice.
3. Pass the path to the dataset to the TISSUENET_DIR environment variable.

split	num_images	image_size
train	2580	512
val	3118	512
test	1324	256

Summary of the benchmarks:

RheaPose Performance

split	time(s)	loss
train	2.74e+02	5.42e-05
val	7.28e+01	5.70e-05
test	3.64e+01	7.63e-05

Cellpose

split	time(s)	loss
train	4.32e+03	3.57e-01
val	2.06e+03	4.25e-01
test	8.28e+02	3.96e-01

RheaPose Multi-Trip

split	num_trips	losses
train	10	[(8.62e-05, 8.62e-05), (0.00e+00, 8.62e-05), (0.00e+00, 8.62e-05), (0.00e+00, 8.62e-05), (0.00e+00, 8.62e-05), (0.00e+00, 8.62e-05), (0.00e+00, 8.62e-05), (0.00e+00, 8.62e-05), (0.00e+00, 8.62e-05), (0.00e+00, 8.62e-05)]
val	10	[(7.63e-05, 7.63e-05), (0.00e+00, 7.63e-05), (0.00e+00, 7.63e-05), (0.00e+00, 7.63e-05), (0.00e+00, 7.63e-05), (0.00e+00, 7.63e-05), (0.00e+00, 7.63e-05), (0.00e+00, 7.63e-05), (0.00e+00, 7.63e-05), (0.00e+00, 7.63e-05)]
test	10	[(5.65e-05, 5.65e-05), (0.00e+00, 5.65e-05), (0.00e+00, 5.65e-05), (0.00e+00, 5.65e-05), (0.00e+00, 5.65e-05), (0.00e+00, 5.65e-05), (0.00e+00, 5.65e-05), (0.00e+00, 5.65e-05), (0.00e+00, 5.65e-05), (0.00e+00, 5.65e-05)]

Cellpose Multi-Trip

split	num_trips	losses
train	10	[(3.85e-01, 3.85e-01), (0.00e+00, 3.85e-01), (0.00e+00, 3.85e-01), (0.00e+00, 3.85e-01), (0.00e+00, 3.85e-01), (0.00e+00, 3.85e-01), (0.00e+00, 3.85e-01), (0.00e+00, 3.85e-01), (0.00e+00, 3.85e-01), (0.00e+00, 3.85e-01)]
val	10	[(4.12e-01, 4.12e-01), (0.00e+00, 4.12e-01), (0.00e+00, 4.12e-01), (0.00e+00, 4.12e-01), (0.00e+00, 4.12e-01), (0.00e+00, 4.12e-01), (0.00e+00, 4.12e-01), (0.00e+00, 4.12e-01), (0.00e+00, 4.12e-01), (0.00e+00, 4.12e-01)]
test	10	[(3.78e-01, 3.78e-01), (0.00e+00, 3.78e-01), (0.00e+00, 3.78e-01), (0.00e+00, 3.78e-01), (0.00e+00, 3.78e-01), (0.00e+00, 3.78e-01), (0.00e+00, 3.78e-01), (0.00e+00, 3.78e-01), (0.00e+00, 3.78e-01), (0.00e+00, 3.78e-01)]