/Pedestrian-Detection

Moving from Non-Neural to Neural methods for Object detection (HoG-SVM to Faster RCNN)

Primary LanguagePython

Pedestrian Detection

In this work we show the transition from non-neural methods, like Histogram-of-Gradients + SVM, to neural methods, like Faster RCNN, for object detection, specifically, pedestrian detection. We use Penn-Fudan Pedestrian Detection Dataset for evaluating our model's performance.

Results

Model Mean Average Precision (mAP) Average Recall @ 1dpi Average Recall @ 10dpi
Pretrained HoG Detector (on INRIA Person dataset) 0.04 0.06 0.14
Custom HoG Detector 0.15 0.13 0.29
Faster-RCNN 0.76 0.30 0.82

Installation

git clone https://github.com/sm354/Pedestrian-Detection.git
cd Pedestrian-Detection
pip install -r requirements.txt

PennFudanPed_train.json, and PennFudanPed_val.json contains COCO annotations for a randomly generated train-val split of the PennFudan dataset.

Download Penn-Fudan Dataset
wget https://www.cis.upenn.edu/~jshi/ped_html/PennFudanPed.zip
unzip PennFudanPed.zip
Download SVM model weights
gdown 1zfU44JxyHCUSWJ7ngiKyqJocgdF82pVt

Running Models

1. Pretrained HoG Detector

python eval_hog_pretrained.py --root <path to dataset root directory> --test <path to test json> --out <path to output json>

2. Custom HoG Detector

Training (HoG descriptors + SVM Model)

python train_hog_custom.py --root <path to dataset root directory> --train <path to train json> --model <path to save trained SVM model>

Testing

python eval_hog_custom.py --root <path to dataset root directory> --test <path to test json> --out <path to output json> --model <path to trained SVM model>

3. Faster RCNN

python eval_faster_rcnn.py --root <path to dataset root directory> --test <path to test json> --out <path to output json>

Evaluation script

python eval_detections.py --gt <path to ground truth annotations json> --pred <path to detections json>

The script eval_detections.py takes in ground truth annotations and predicted detections for the evaluation dataset and computes the following metrics:

  1. Average Precision, computed over 10 IOU thresholds in the range 0.5:0.05:0.95
  2. Average Recall computed at 1 detection per image.
  3. Average Recall comptued at 10 detections per image.

Authors

Course assignment in Computer Vision course (course webpage) taken by Prof. Chetan Arora