FLIR Dataset

Dataset Overview

Dataset Name

FLIR Dataset

Dataset Link

https://www.flir.in/oem/adas/adas-dataset-form/

Dataset Content

• Type: Primarily three types of thermal‑imaging images.
• Details:
  • train: 8,862 thermal images.
  • val: 1,366 thermal images.
  • video: 4,224 thermal images.

Dataset Usage

• Training & Validation: Images in the "train" and "val" folders are used respectively for training and validating a YOLOv3 detector.
• Object Tracking: The "video" folder is used for tracking detected objects.

Dataset Performance Metrics

• mAP (Mean Average Precision):
  • YOLOv3 @ 416:
    • Overall mAP: 0.38033
    • Main three‑class mAP: 0.50563
  • YOLOv3 @ 608:
    • Overall mAP: 0.44037
    • Main three‑class mAP: 0.58576

Dataset Limitations

• Low AP for Dog Class: Due to the low occurrence of this class in the dataset, the AP is relatively low.

Dataset Application Example

• Output Video: Tracking results can be viewed at the following link:
  • Link: https://drive.google.com/file/d/1ICfcUHrX_Cd3yBGXYq2EE8HUF4-W94Rp/view?usp=sharing

Model Weights

• YOLOv3 @ 608:
  • Link: https://drive.google.com/file/d/19x6B6oYB8Y4o3Sqp5YXAPBEI38yP2Uq9/view?usp=sharing
• YOLOv3 @ 416:
  • Link: https://drive.google.com/file/d/1M2xr2aW0Zj1ghM91FF4EoBgwCQpRNz2p/view?usp=sharing

Running Code Guide

• Step 1: Run the "YOLOv3 detector" to save detection results to a text file.
• Step 2: Use the "SORT tracker" code to process the stored detection results and save images.
• Step 3: Create a video from the images with tracking overlays.