B-FPGM: Lightweight Face Detection via Bayesian-Optimized Soft FPGM Pruning
- 1. B-FPGM: Lightweight Face Detection via Bayesian-Optimized Soft FPGM Pruning Nikolaos Kaparinos, Vasileios Mezaris CERTH-ITI, Thermi, Thessaloniki, Greece Real-World Surveillance Workshop @ WACV 2025
- 2. The Growing Demand for Compact AI Models ● The deployment of AI models on mobile devices, such as smartphones and drones, is increasingly common. ● Thus, the need for compact and efficient AI models has dramatically increased. ● Face detectors are a type of model commonly deployed on mobile devices. ● Lightweight face detectors have been proposed in the literature. ● They utilize lightweight backbone networks and other optimization techniques, such as pruning. 2
- 3. Network Pruning ● Network pruning is a technique used to reduce the number of parameters in a model. ● Pruning methods can also be classified into uniform and non-uniform approaches. ● FPGM pruning is a structured pruning approach that has demonstrated high performance in the literature. ● Soft Filter Pruning (SFP) is a pruning method that allows the pruned filters to be updated during subsequent training steps. 3
- 4. B-FPGM ● This work proposes, B-FPGM, a novel non-uniform face detection network pruning technique. ● This work represents the first application of Bayesian optimization to structured pruning as well as non-uniform pruning in the literature. ● B-FPGM divides the network layers into 6 groups and employes Bayesian optimization to optimize the pruning rate of each group. ● The optimal pruning rates are then applied alongside FPGM pruning and SFP. 4
- 5. B-FPGM Advantages ● B-FPGM offers flexibility through its non-universal pruning approach. ● It eliminates the need for engineering expertise to define rules for optimal pruning rates, effectively taking the ‘human out of the loop’. ● At the same time, it avoids utilizing Reinforcement Learning, which comes with significant drawbacks. 5
- 7. Bayesian optimization step ● The Bayesian optimization step is employed to identify the optimal pruning rate for each layer group, given a target overall sparsity. ● In each iteration, the pre-trained network is soft-pruned and trained for one epoch. ● The objective function value is equal to the validation loss, plus an additional term to ensure that the network is pruned approximately at the target overall sparsity. 7
- 8. Network Layer Groups 8 The number of parameters in each network layer group. EResFD model architecture and layer groups.
- 10. Experimental Setup ● All our experiments were applied to EResFD, the currently smallest (in number of parameters) well-performing face detector of the literature. ● A small ablation study with a second small face detector, EXTD, is also reported. ● The experiments were performed using the WIDER FACE dataset. ○ 12941 training images ○ Three validation subsets based on difficulty: Easy (1146 images), Medium (1079 images), Hard (1001 images) ● Experiments were conducted with target pruning rates ranging from 10% to 60%. 10
- 11. Results on EResFD using the WIDER FACE dataset 11 Hard Subset Group pruning rates determined by Bayesian optimization. T is the target pruning rate. 10% 10% 20% 20% 30% 30% 40% 40% 50% 50% 60% 60%
- 12. Comparison with SoA models 12
- 13. Robustness to Randomness 13 Mean mAP ± standard deviation of B-FPGM on EResFD across five runs, using different random seeds, for 20% target pruning rate.
- 14. Number of layer groups ablation 14 MAP of B-FPGM on EResFD, on WIDER FACE (Easy, Medium, Hard subsets), for different network layer groupings. N is the number of layer groups and T is the target pruning rate.
- 15. Inference visual example 15 EResFD 50% pruned using B-FPGM
- 16. Thank you for your attention! Questions? Nikolaos Kaparinos, kaparinos@iti.gr Vasileios Mezaris, bmezaris@iti.gr Source code and pruned models available at: https://github.com/IDT-ITI/B-FPGM This work was supported by the EU Horizon Europe and Horizon 2020 programmes under grant agreements 101070093 vera.ai and 951911 AI4Media, respectively. 16