Skip to content

Latest commit

 

History

History

Pedestron

README.md

BlockCopy - Pedestrian detection with Pedestron

This codebase for pedestrian detection is based on Pedestron. We removed architectures and datasets that are not useful for our experiments for clarity.

Installation

Note: a Dockerfile (contributed by a user, thanks!) is available in ./docker/Dockerfile

Requirements

We tested with

  • Pytorch 1.9.1
  • CUDA 11.4
  • CuPy 9.5
  • GCC 9.3
  • mmdet 0.6.0

Installation

  • Install BlockCopy, its requirements and the cityscapes dataset as specified in the BlockCopy readme
  • Install cython: pip install cython
  • Install pedestron: python setup.py develop (if CUDA extensions do not build, check your CUDA setup and if the correct paths are set)

Dataset

Prepare Cityscapes video data as in the main BlockCopy readme and update the img_root in the config files located in

configs/elephant/cityperson

Model checkpoints

Get the CSP model from Pedestron from Google drive. Place in ./checkpoints/ dir resulting in the following structure: ./checkpoints/csp/epoch_72.pth

Note that you have to rename the file (remove .stu extension)

Testing

Dynamic BlockCopy model

python ./tools/test_city_person.py configs/elephant/cityperson/csp_r50_clip_blockcopy_030.py ./checkpoints/csp/epoch_ 72 73 --out results/csp_blockcopy_t030.json --num-clips-warmup 400 --num-clips-eval -1

Resulting in

Computational cost (avg per img): 380.097 GMACs over 10000 images
======= FLOPSCOUNTER =======
batches: 10000
# depth 0: 
model                (MMDataParallel):      380.1 GMac
# depth 1: 
module               (CSPBlockCopy):      380.1 GMac
# depth 2: 
backbone             (ResNet    ):      85.71 GMac
bbox_head            (CSPHead   ):     231.09 GMac
neck                 (CSPNeck   ):      56.79 GMac
policy               (PolicyTrainRL):       6.51 GMac

Checkpoint 72: [Reasonable: 11.44%], [Reasonable_Small: 15.31%], [Heavy: 40.56%], [All: 37.47%]

With visualisations of detections, executed blocks, information gain (written to output/csp_blockcopy_t030):

python ./tools/test_city_person.py configs/elephant/cityperson/csp_r50_clip_blockcopy_030.py ./checkpoints/csp/epoch_ 72 73 --out results/csp_blockcopy_t030.json  --save_img --save_img_dir output/csp_blockcopy_t030 --num-clips-warmup 400 --num-clips-eval -1

Static standard model

python ./tools/test_city_person.py configs/elephant/cityperson/csp_r50_clip.py ./checkpoints/csp/epoch_ 72 73 --out results/csp_blockcopy_t050.json --num-clips-warmup 400 --num-clips-eval -1