MultiPathNet is a Torch-7 implementation of the “A MultiPath Network for Object Detection” paper (BMVC 2016), developed by Facebook AI Research. It extends the Fast R-CNN framework by introducing multiple network “paths” to enhance feature extraction and object recognition robustness. The MultiPath architecture incorporates skip connections and multi-scale processing to capture both fine-grained details and high-level context within a single detection pipeline. This results in improved detection accuracy across various object sizes and categories compared to standard single-path architectures. The repository supports training, evaluation, and visualization for object detection tasks on popular datasets such as PASCAL VOC and MS COCO. It provides pre-trained models for VGG, AlexNet, and ResNet backbones, along with integration for SharpMask and DeepMask proposal generators.
Features
- Implements Fast R-CNN and MultiPath object detection in Torch-7
- Multi-scale and multi-path feature extraction for improved detection accuracy
- Supports major backbones: AlexNet, VGG, ResNet, and Inception-v3
- Integrated with SharpMask and DeepMask for region proposal generation
- Multi-GPU and EC2-ready training with data parallelism
- Pre-trained models and scripts for COCO and PASCAL VOC evaluation
Project Samples
