PSXIII-B-6 - Effect of Camera Exposure Time on Image Segmentation and Body Weight Prediction

Abstract: External conditions such as lightness and brightness in a scene might have a tremendous effect on the quality of image acquisition, and this may influence on the prediction of phenotypes of interest. Here we investigated the impact of camera exposure time on the performance of image segmentation and body weight (BW) prediction. Top view depth images were acquired from 84 dairy cows using exposure time automatically defined by the camera (Exp0) given the scene light conditions, and manually set as 1,000 milliseconds (Exp1) and 4,000 milliseconds (Exp4) for all cows. Three images per cow from each experiment (n = 252) had the animal body manually traced. A 10-fold cross-validation (random by cow) strategy was used to train and validate three U-Net networks, one per experiment, for semantic segmentation. The trained segmenters were tested on each experiment dataset using intersection over union (IoU) as an evaluation metric. Biological features (BF) including area, volume, circularity, and major axis length were extracted from the segmented images. BF were used as predictors in linear models to predict the BW, and the models’ performance were evaluated using coefficient of determination (R²). IoU remained constant (0.98) when training U-Net using images from Exp0, Exp1, and Exp4 and testing on Exp0 and Exp1, and decreased (0.92) when testing on Exp4. R² of 0.79, 0.78, and 0.08 were observed to predict BW when the model was trained using BF from Exp0 and Exp1 and tested using BF from Exp0, Exp1, and Exp4, respectively. Lower R² (0.33, 0.34, and 0.25, respectively) were observed using BF from Exp4 to train the model and test using BF from Exp0, Exp1, and Exp4. Overall, higher exposure times (Exp4) resulted in poorer segmentation quality and consequently less precise BW predictions.