Object Detection · Pose Estimation · Stereo Vision

AI powered Fish Size Measurement

We developed a computer vision system to measure fish size from stereo camera data. Our method achieves accurate size estimation while being efficient enough for real-time deployment in aquaculture settings.

Dataset Fish detection and pose estimation dataset collected from stereo cameras in aquaculture environments.

Framework PyTorch · MMDetection · MMPose

Hardware A100 GPU

Task Object Detection · Pose Estimation · Stereo Matching

Problem Statement

Fish size measurement is critical for aquaculture management, but traditional methods are labor-intensive and error-prone. The size of fish can act as a proxy for health and growth, and accurate measurements can help optimize feeding schedules.

Method Overview

Faster R-CNN based object detection for fish localization, followed by a top-down pose estimation module to identify keypoints on the fish body using HR-Net and Hourglass models.

Visual Results

Fish association across frames

Image 2

Ground Truth Comparison

Image 3

Failure Case

Image 4

Feature Visualization

Quantitative Results - Object Detection

mAP values for object detection methods on our dataset with Faster R-CNN-R50 and R101 backbones.

Method	mAP	AP50	Image Size
Faster R-CNN-R50	77.20	85.80	1280x720
Faster R-CNN-R101	78.50	86.00	1280x720

Quantitative Results - Pose Estimation (top-down)

AP values for pose estimation methods on our dataset with HRNet-32 and Hourglass-52 backbones.

Method	AP	AP75	Image Size
HRNet-32	75.31	94.77	384x288
Hourglass-52	65.40	81.71	256x256

Quantitative Results - Fish association for Stereo Matching

Precision and Recall values for fish association across left and right frames on our dataset.

Method	Precision	Recall
Best-Match	81.36	77.38
Minimum Cost Assignment	90.19	82.31

Quantitative Results - Length measurement Errors

Length measurement errors for different methods of depth estimation at keypoints. We use ground-truth depth and estimated depths at the keypoints to calculate the errors.

Method	Error (cm)
Predicted Keypoints + GT Depth	1.2
Predicted Keypoints + Estimated Depth	17.5
Predicted Keypoints + Estimated Depth (Disparity >= 50px)	1.6

Underwater object detection and pose estimation.
Keypoints from pose estimation used for stereo matching.
Depth estimation at keypoints and length measurement.