Perceptual Piercing: Human Visual Cue-Based Object Detection in Low Visibility Conditions

A novel deep-learning framework, inspired by the atmospheric scattering model and human visual cortex mechanisms, to enhance object detection under poor visibility scenarios such as fog, smoke, and haze. Implemented a multi-tiered strategy based on human-like visual cues that integrates an initial quick detection, followed by target-specific dehazing, and concludes with an in-depth detection phase.

Installation

To set up the repository, follow these steps:

1. Clone the repository

git clone https://github.com/ashu1069/perceptual-piercing.git
cd perceptual-piercing

2. Create a Virtual Environment

python3 -m venv venv
source venv/bin/activate

3. Install Required Libraries

pip install -r requirements.txt

4. Download pre-trained models: Download the YOLOv5 and YOLOv8 pre-trained models and place them in the models/ directory:

AOD-NetX

1. Dataset Preparation: Prepare datasets using the corresponding dataloader.py script in the AOD-NetX/ directory.
2. Generate Bounding Box Annotations: Use yolov5_annotations.py and yolov8_annotations.py scripts to create bounding box annotations from the dataset directory.
3. Model Training: Run train.py in the AOD-NetX/ directory with clean images, foggy images, and the generated bounding box directory.
4. Model Testing: Execute test.py in the AOD-NetX/ directory to assess the dehazing performance of the model.

Full Pipeline Evaluation

To perform a complete evaluation of the pipeline, follow these steps:

• Quick Object Detection with Lightweight YOLO Models:

  • Use the yolovx_annotations.py script to perform initial object detection on the input images.
  • This step generates preliminary bounding boxes using a lightweight YOLO model (YOLOv5n or YOLOv8n).

• Dehazing the Detected Regions:

  • Use the bounding boxes from Step 1 to focus dehazing on specific regions of the images.
  • Run test_FC.py if evaluating on the Foggy Cityscapes dataset, or test_RESIDE.py for the OTS and RTTS subsets of the RESIDE dataset.

• Final Object Detection with High-Precision YOLO Models:

  • After dehazing, perform a final object detection using high-precision YOLO models (YOLOv5x or YOLOv8x) by re-running yolovx_annotations.py on the dehazed outputs.

• Performance Evaluation:

  • Run metrics.py to evaluate the detection performance using standard metrics such as mean Average Precision (mAP).
  • This script compares the detection results on dehazed images with ground truth annotations.

Evaluation Scripts

1. General Dehazing Evaluation: Run dehazing_evaluation.py in the evaluation/ directory to evaluate the performance of any general dehazing method.
2. Foggy Cityscapes Evaluation: Use test_FC.py in the evaluation/ directory to measure dehazing performance on the Foggy Cityscapes dataset.
3. RESIDE Dataset Evaluation: Use test_RESIDE.py in the evaluation/ directory to evaluate dehazing performance on the OTS and RTTS subsets of the RESIDE dataset.
4. Object Detection Performance: Execute metrics.py in the evaluation/ directory to assess object detection performance using the dehazed outputs.