This project performs inference of baseline detection methods on datasets, both synthetic and real, and performs a sensitivity analysis on the results.
Given input detection models (RGB --> (x0,y0,x1,y1) it aims at :
- Analyzing their sensitivity toward parameters such as
- frame parameters (luminosity, weather, camera angle, camera distance)
- boundig box parameters (height, occlusion)
- Provide an exhaustive benchmark on MoTSynth and EuroCityPerson datasets to make a proof of concept of the method, yielding empirical results indicating what works and what does not work in the synthetic dataset.
Assuming each twincity folder is stored in /home/raphael/work/datasets/twincity-Unreal/v5 :
Run python main.py -d twincity -r /home/raphael/work/datasets/twincity-Unreal/v5 --max_samples 50 -output results_new -frame -gt --plot_image
- Download motsynth
- Download eurocityperson (research purpose only)
For more information on EuroCityPerson and its license see https://eurocity-dataset.tudelft.nl/
For more information on MoTSynth and its license see https://aimagelab.ing.unimore.it/imagelab/page.asp?IdPage=42
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch torchmetrics tqdm pandas seaborn conda install -c conda-forge opencv This post did help https://discuss.pytorch.org/t/userwarning-cuda-initialization-cuda-unknown-error-this-may-be-due-to-an-incorrectly-set-up-environment-e-g-changing-env-variable-cuda-visible-devices-after-program-start-setting-the-available-devices-to-be-zero/129335/2 scikit-learn cpu conda install pytorch torchvision torchaudio -c pytorch See mmdet / mmcv for install
- PennFudan
- Isgroup : ignore region ? (if not it generates false positives because good detector find each people in the group instead of just 1)
- MoTSynth
- ??
- ECP
- Twincity
- ??
- Twincity
- Hendled Glitch via ignore regions : too big bbox or multiple colors
typical / atypical above 50pix or not ?