Hybrid Faster R-CNN — Learnable Weighted Fusion
Model
Hybrid Faster R-CNN object detector combining:
The two backbone branches are combined using learnable weighted
feature fusion.
The fusion weights are initialized at 0.5 / 0.5 and learned during
training.
Architecture
Input Image
→ ResNet-101
→ Inception-V3
→ Learnable Weighted Fusion
→ 256-channel feature map
→ Faster R-CNN
→ Object Detection
Detector
- Framework: Detectron2
- Detector: Faster R-CNN
- Classes: 11
- Feature map: fused
- Channels: 256
- Stride: 32
- Training iterations: 30,000
Validation Results
AP: 51.3646
AP50: 84.2608
AP75: 57.2804
APm: 28.2779
APl: 52.4605
AR@100: 58.9
Baseline
The baseline model uses feature concatenation.
Baseline:
AP: 52.6155
AP50: 87.8719
AP75: 58.2183
The learnable weighted fusion therefore provides a closely comparable
alternative to simple feature concatenation.
Purpose
This model represents the learnable weighted-fusion ablation used to
investigate the effect of feature-fusion mechanisms in the hybrid
ResNet-101 + Inception-V3 detector.
The original concatenation baseline is preserved separately.
