How ubenz55555/mscnn works
This repository is an MS-CNN object-detection implementation built on Caffe. The documented workflow trains a unified multi-scale network, combining proposal and detection subnetworks, then runs KITTI detection/proposal inference and saves results for evaluation. The graph also shows Caffe’s model execution and optimization layers and Python-facing utilities. README documents product capabilities, but the referenced MS-CNN example and evaluation files are absent from the supplied tree; those links are therefore documented rather than source-verified.
Main entry points: net.cpp Network Graph · layer.hpp Layer Operations · detector.py Python Detection · solver.cpp Solver
Components 20
MS-CNN Detection
- Unified MS-CNN
- Object Proposals
src/caffe/layers/proposal_target_layer.cpp - Object Detection
src/caffe/layers/detection_loss_layer.cpp - Detection Metrics
src/caffe/layers/detection_accuracy_layer.cpp - Image GT Loader
src/caffe/layers/image_gt_data_layer.cpp - Detection Results
- KITTI Evaluation
Caffe Runtime
- Network Graph
src/caffe/net.cpp - Layer Operations
include/caffe/layer.hpp - Layer Factory
include/caffe/layer_factory.hpp - Tensor Blobs
include/caffe/blob.hpp - Compute Kernels
include/caffe/util/math_functions.hpp
Training
- Solver
src/caffe/solver.cpp - SGD Solvers
src/caffe/solvers/sgd_solver.cpp
User Interfaces
- Python API
python/caffe/_caffe.cpp - Python Image IO
python/caffe/io.py - Python Detection
python/caffe/detector.py - MATLAB Wrapper
matlab/+caffe/private/caffe_.cpp
Other components
- Engineer
- KITTI Data
Connections 23
- Engineer → KITTI Data: supplies data
- KITTI Data → Image GT Loader: provides images
- Image GT Loader → Unified MS-CNN: feeds training
- Unified MS-CNN → Object Proposals: includes proposal stage
- Unified MS-CNN → Object Detection: includes detection stage
- Engineer → Solver: starts training
- Solver → Network Graph: optimizes network
- Solver → SGD Solvers: selects solver
- Network Graph → Layer Factory: creates layers
src/caffe/net.cpp - Layer Factory → Layer Operations: constructs
include/caffe/layer_factory.hpp - Network Graph → Layer Operations: sets up layers
src/caffe/net.cpp - Layer Operations → Tensor Blobs: operates on tensors
include/caffe/layer.hpp - Layer Operations → Compute Kernels: uses kernels
include/caffe/layer.hpp - Engineer → Python API: uses interface
python/caffe/_caffe.cpp - Python API → Network Graph: exposes network
python/caffe/_caffe.cpp - Engineer → MATLAB Wrapper: uses interface
- Python Image IO → Python Detection: prepares images
- Python Detection → Python API: runs inference
- Unified MS-CNN → Detection Metrics: evaluates detections
- Engineer → Python Detection: runs demo
- Python Detection → Detection Results: saves results
- Detection Results → KITTI Evaluation: formats for KITTI
- KITTI Evaluation → Engineer: reports performance