Huaqing Yu-1,2, Yi He-1, Peng Du-1, Lu Song-11- Tianjin Key Laboratory of Intelligent Unmanned Swarm Technology and System,School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China2- Tencent, Shenzhen 518000, China{huatsing, heyi, dupeng2022, songlu}@tju.edu.cnAbstract:Face images contain a wide variety of attribute information. In this paper, we propose a generalized framework for joint estimation of ordinal and nominal attributes based on information sharing. We tackle the correlation problem between heterogeneous attributes using hard parameter sharing of shallow features, and trade-off multiple loss functions by considering homoskedastic uncertainty for each attribute estimation task. This leads to optimal estimation of multiple attributes of the face and reduces the training cost of multitask learning. Experimental results on benchmarks with multiple face attributes show that the proposed approach has superior performance compared to state of the art. Finally, we discuss the bias issues arising from the proposed approach in face attribute estimation and validate its feasibility on edge systems.
Race Single Task ResNet AlexNet DenseNet VGGNet Jamoliddin et al., Proposed 60.17 58.6 55.96 59.22 57.06 61.34 64.74 90.09 87.4 85.1 87.28 83.63 88.67 90.91 76.23 79.98 1results are reported using the accuracy with age group. denotes unavailable metrics in related literature. 3) Comparison of estimation error in ordinal attribute tasks: For ordinal attribute estimation task, we computed the Mean Square Error (MSE) and Mean Absolute Error (MAE) for the age estimation task in Table III. The proposed approach exhibits lower estimation errors compared to other approaches due to the improved generalization performance induced by DMTL. B. Ablation Study As shown in Figure 3, we depict the variation in weights of task losses during the training process using homoscedastic uncertainty, where i is normalized using Equation 16. 1 (cid:80)M
id: 02f581f7b0253f7c24d3480be1333e9f - page: 4
It can be observed that on the Adience dataset, the weight for the age estimation task age is comparable to that of the gender recognition task gender, while on the UTKFace dataset, the weight for the age estimation task age is very low. To ensure that tasks with different magnitudes of loss values are adequately trained, Equation 2 or Equation 15 assigns smaller weights to tasks with larger losses. We present the results of ablation experiments in Table IV. In the Proposed without ordinal opti approach, we neglected the special treatment for ordinal tasks, while in the Proposed 4 (16) XXX, VOL. XX, NO. X, XX XXXX TABLE III COMPARING THE ESTIMATION ERROR OF DIFFERENT APPROACHES IN AGE ESTIMATION TASK Approach Adience MSE MAE MSE UTKFace
id: e374e33ea7c43d79f379c70aeaf1978a - page: 4
MAE Single Task Li et al., MobileNet v2 Sumit Kothari FaceNet Proposed 0.9266 0.2409 0.4194 0.41 0.37 0.1639 116.87 66.3958 6.73 7.29 6.76 6.71 5.323 denotes unavailable metrics in related literature. (a) (b) Fig. 3. Convergence curves of task weights training on two datasets, i represents the weight of the attribute i estimation task in the Joint loss L. (a) Adience. (b) UTKFace. without uncertainty approach, we selected the approximate optimal weights for multi-task learning through manual adjustments. For ordinal task, we computed the Cumulative Score (CS) for three approaches on the UTKFace benchmark and illustrated it in Figure 4. This will help us to validate different approaches in ordinal attribute prediction performance. The CS is calculated as follows:
id: 443b781517f2bc4e3e56976874959e42 - page: 5
Ni N where N is the total number of test images, and Ni is the number of test images whose the absolute error between the estimated value and the label is less than order i. CS(i) = 100% Fig. 4. The comparison of age estimation with CS metric on UTKFace benchmark. Through ablation experiments, we demonstrate the effectiveness of reducing estimation errors by transforming ordinal task into a linear combination of binary classifications subproblems. Furthermore, adjusting the weights of multi-task loss through uncertainty is beneficial.
id: 78424645d928568c8934479fd3b8e68b - page: 5