Optimize dbnet (#464)
* Fix: syntax error in common.hpp , cuda & tensorrt not included in CMakeLists.txt * Fix: layer warining while building in common.hpp and denet.cpp * Fix: NOT droping mini boxes and low score boxes. * Fix: The prediction boxes are NOT expanded by the specified proportion, which will lead to serious errors
This commit is contained in:
PeterH0323
GitHub
parent
5b90354d28
commit
b74db15d87
@ -1,6 +1,6 @@
|
||||
cmake_minimum_required(VERSION 2.6)
|
||||
|
||||
project(yolov4)
|
||||
project(dbnet)
|
||||
|
||||
add_definitions(-std=c++11)
|
||||
|
||||
@ -12,10 +12,24 @@ find_package(CUDA REQUIRED)
|
||||
|
||||
include_directories(${PROJECT_SOURCE_DIR}/include)
|
||||
|
||||
# cuda
|
||||
include_directories(/usr/local/cuda/include)
|
||||
link_directories(/usr/local/cuda/lib64)
|
||||
# tensorrt
|
||||
include_directories(/usr/include/x86_64-linux-gnu/)
|
||||
link_directories(/usr/lib/x86_64-linux-gnu/)
|
||||
|
||||
find_package(OpenCV)
|
||||
include_directories(${OpenCV_INCLUDE_DIRS})
|
||||
|
||||
add_executable(dbnet ${PROJECT_SOURCE_DIR}/dbnet.cpp)
|
||||
aux_source_directory(. DIRSRCS)
|
||||
|
||||
# clipper
|
||||
include_directories(./ ./clipper)
|
||||
add_subdirectory(clipper)
|
||||
|
||||
add_executable(dbnet ${DIRSRCS})
|
||||
target_link_libraries(dbnet clipper)
|
||||
target_link_libraries(dbnet nvinfer)
|
||||
target_link_libraries(dbnet cudart)
|
||||
target_link_libraries(dbnet ${OpenCV_LIBS})
|
||||
|
||||
4
dbnet/clipper/CMakeLists.txt
Normal file
4
dbnet/clipper/CMakeLists.txt
Normal file
@ -0,0 +1,4 @@
|
||||
cmake_minimum_required(VERSION 2.6)
|
||||
|
||||
aux_source_directory(. DIR_CLIPPER_SRCS)
|
||||
add_library(clipper ${DIR_CLIPPER_SRCS})
|
||||
4629
dbnet/clipper/clipper.cpp
Normal file
4629
dbnet/clipper/clipper.cpp
Normal file
File diff suppressed because it is too large
Load Diff
406
dbnet/clipper/clipper.hpp
Normal file
406
dbnet/clipper/clipper.hpp
Normal file
@ -0,0 +1,406 @@
|
||||
/*******************************************************************************
|
||||
* *
|
||||
* Author : Angus Johnson *
|
||||
* Version : 6.4.2 *
|
||||
* Date : 27 February 2017 *
|
||||
* Website : http://www.angusj.com *
|
||||
* Copyright : Angus Johnson 2010-2017 *
|
||||
* *
|
||||
* License: *
|
||||
* Use, modification & distribution is subject to Boost Software License Ver 1. *
|
||||
* http://www.boost.org/LICENSE_1_0.txt *
|
||||
* *
|
||||
* Attributions: *
|
||||
* The code in this library is an extension of Bala Vatti's clipping algorithm: *
|
||||
* "A generic solution to polygon clipping" *
|
||||
* Communications of the ACM, Vol 35, Issue 7 (July 1992) pp 56-63. *
|
||||
* http://portal.acm.org/citation.cfm?id=129906 *
|
||||
* *
|
||||
* Computer graphics and geometric modeling: implementation and algorithms *
|
||||
* By Max K. Agoston *
|
||||
* Springer; 1 edition (January 4, 2005) *
|
||||
* http://books.google.com/books?q=vatti+clipping+agoston *
|
||||
* *
|
||||
* See also: *
|
||||
* "Polygon Offsetting by Computing Winding Numbers" *
|
||||
* Paper no. DETC2005-85513 pp. 565-575 *
|
||||
* ASME 2005 International Design Engineering Technical Conferences *
|
||||
* and Computers and Information in Engineering Conference (IDETC/CIE2005) *
|
||||
* September 24-28, 2005 , Long Beach, California, USA *
|
||||
* http://www.me.berkeley.edu/~mcmains/pubs/DAC05OffsetPolygon.pdf *
|
||||
* *
|
||||
*******************************************************************************/
|
||||
|
||||
#ifndef clipper_hpp
|
||||
#define clipper_hpp
|
||||
|
||||
#define CLIPPER_VERSION "6.4.2"
|
||||
|
||||
//use_int32: When enabled 32bit ints are used instead of 64bit ints. This
|
||||
//improve performance but coordinate values are limited to the range +/- 46340
|
||||
//#define use_int32
|
||||
|
||||
//use_xyz: adds a Z member to IntPoint. Adds a minor cost to perfomance.
|
||||
//#define use_xyz
|
||||
|
||||
//use_lines: Enables line clipping. Adds a very minor cost to performance.
|
||||
#define use_lines
|
||||
|
||||
//use_deprecated: Enables temporary support for the obsolete functions
|
||||
//#define use_deprecated
|
||||
|
||||
#include <vector>
|
||||
#include <list>
|
||||
#include <set>
|
||||
#include <stdexcept>
|
||||
#include <cstring>
|
||||
#include <cstdlib>
|
||||
#include <ostream>
|
||||
#include <functional>
|
||||
#include <queue>
|
||||
|
||||
namespace ClipperLib {
|
||||
|
||||
enum ClipType { ctIntersection, ctUnion, ctDifference, ctXor };
|
||||
enum PolyType { ptSubject, ptClip };
|
||||
//By far the most widely used winding rules for polygon filling are
|
||||
//EvenOdd & NonZero (GDI, GDI+, XLib, OpenGL, Cairo, AGG, Quartz, SVG, Gr32)
|
||||
//Others rules include Positive, Negative and ABS_GTR_EQ_TWO (only in OpenGL)
|
||||
//see http://glprogramming.com/red/chapter11.html
|
||||
enum PolyFillType { pftEvenOdd, pftNonZero, pftPositive, pftNegative };
|
||||
|
||||
#ifdef use_int32
|
||||
typedef int cInt;
|
||||
static cInt const loRange = 0x7FFF;
|
||||
static cInt const hiRange = 0x7FFF;
|
||||
#else
|
||||
typedef signed long long cInt;
|
||||
static cInt const loRange = 0x3FFFFFFF;
|
||||
static cInt const hiRange = 0x3FFFFFFFFFFFFFFFLL;
|
||||
typedef signed long long long64; //used by Int128 class
|
||||
typedef unsigned long long ulong64;
|
||||
|
||||
#endif
|
||||
|
||||
struct IntPoint {
|
||||
cInt X;
|
||||
cInt Y;
|
||||
#ifdef use_xyz
|
||||
cInt Z;
|
||||
IntPoint(cInt x = 0, cInt y = 0, cInt z = 0): X(x), Y(y), Z(z) {};
|
||||
#else
|
||||
IntPoint(cInt x = 0, cInt y = 0): X(x), Y(y) {};
|
||||
#endif
|
||||
|
||||
friend inline bool operator== (const IntPoint& a, const IntPoint& b)
|
||||
{
|
||||
return a.X == b.X && a.Y == b.Y;
|
||||
}
|
||||
friend inline bool operator!= (const IntPoint& a, const IntPoint& b)
|
||||
{
|
||||
return a.X != b.X || a.Y != b.Y;
|
||||
}
|
||||
};
|
||||
//------------------------------------------------------------------------------
|
||||
|
||||
typedef std::vector< IntPoint > Path;
|
||||
typedef std::vector< Path > Paths;
|
||||
|
||||
inline Path& operator <<(Path& poly, const IntPoint& p) {poly.push_back(p); return poly;}
|
||||
inline Paths& operator <<(Paths& polys, const Path& p) {polys.push_back(p); return polys;}
|
||||
|
||||
std::ostream& operator <<(std::ostream &s, const IntPoint &p);
|
||||
std::ostream& operator <<(std::ostream &s, const Path &p);
|
||||
std::ostream& operator <<(std::ostream &s, const Paths &p);
|
||||
|
||||
struct DoublePoint
|
||||
{
|
||||
double X;
|
||||
double Y;
|
||||
DoublePoint(double x = 0, double y = 0) : X(x), Y(y) {}
|
||||
DoublePoint(IntPoint ip) : X((double)ip.X), Y((double)ip.Y) {}
|
||||
};
|
||||
//------------------------------------------------------------------------------
|
||||
|
||||
#ifdef use_xyz
|
||||
typedef void (*ZFillCallback)(IntPoint& e1bot, IntPoint& e1top, IntPoint& e2bot, IntPoint& e2top, IntPoint& pt);
|
||||
#endif
|
||||
|
||||
enum InitOptions {ioReverseSolution = 1, ioStrictlySimple = 2, ioPreserveCollinear = 4};
|
||||
enum JoinType {jtSquare, jtRound, jtMiter};
|
||||
enum EndType {etClosedPolygon, etClosedLine, etOpenButt, etOpenSquare, etOpenRound};
|
||||
|
||||
class PolyNode;
|
||||
typedef std::vector< PolyNode* > PolyNodes;
|
||||
|
||||
class PolyNode
|
||||
{
|
||||
public:
|
||||
PolyNode();
|
||||
virtual ~PolyNode(){};
|
||||
Path Contour;
|
||||
PolyNodes Childs;
|
||||
PolyNode* Parent;
|
||||
PolyNode* GetNext() const;
|
||||
bool IsHole() const;
|
||||
bool IsOpen() const;
|
||||
int ChildCount() const;
|
||||
private:
|
||||
//PolyNode& operator =(PolyNode& other);
|
||||
unsigned Index; //node index in Parent.Childs
|
||||
bool m_IsOpen;
|
||||
JoinType m_jointype;
|
||||
EndType m_endtype;
|
||||
PolyNode* GetNextSiblingUp() const;
|
||||
void AddChild(PolyNode& child);
|
||||
friend class Clipper; //to access Index
|
||||
friend class ClipperOffset;
|
||||
};
|
||||
|
||||
class PolyTree: public PolyNode
|
||||
{
|
||||
public:
|
||||
~PolyTree(){ Clear(); };
|
||||
PolyNode* GetFirst() const;
|
||||
void Clear();
|
||||
int Total() const;
|
||||
private:
|
||||
//PolyTree& operator =(PolyTree& other);
|
||||
PolyNodes AllNodes;
|
||||
friend class Clipper; //to access AllNodes
|
||||
};
|
||||
|
||||
bool Orientation(const Path &poly);
|
||||
double Area(const Path &poly);
|
||||
int PointInPolygon(const IntPoint &pt, const Path &path);
|
||||
|
||||
void SimplifyPolygon(const Path &in_poly, Paths &out_polys, PolyFillType fillType = pftEvenOdd);
|
||||
void SimplifyPolygons(const Paths &in_polys, Paths &out_polys, PolyFillType fillType = pftEvenOdd);
|
||||
void SimplifyPolygons(Paths &polys, PolyFillType fillType = pftEvenOdd);
|
||||
|
||||
void CleanPolygon(const Path& in_poly, Path& out_poly, double distance = 1.415);
|
||||
void CleanPolygon(Path& poly, double distance = 1.415);
|
||||
void CleanPolygons(const Paths& in_polys, Paths& out_polys, double distance = 1.415);
|
||||
void CleanPolygons(Paths& polys, double distance = 1.415);
|
||||
|
||||
void MinkowskiSum(const Path& pattern, const Path& path, Paths& solution, bool pathIsClosed);
|
||||
void MinkowskiSum(const Path& pattern, const Paths& paths, Paths& solution, bool pathIsClosed);
|
||||
void MinkowskiDiff(const Path& poly1, const Path& poly2, Paths& solution);
|
||||
|
||||
void PolyTreeToPaths(const PolyTree& polytree, Paths& paths);
|
||||
void ClosedPathsFromPolyTree(const PolyTree& polytree, Paths& paths);
|
||||
void OpenPathsFromPolyTree(PolyTree& polytree, Paths& paths);
|
||||
|
||||
void ReversePath(Path& p);
|
||||
void ReversePaths(Paths& p);
|
||||
|
||||
struct IntRect { cInt left; cInt top; cInt right; cInt bottom; };
|
||||
|
||||
//enums that are used internally ...
|
||||
enum EdgeSide { esLeft = 1, esRight = 2};
|
||||
|
||||
//forward declarations (for stuff used internally) ...
|
||||
struct TEdge;
|
||||
struct IntersectNode;
|
||||
struct LocalMinimum;
|
||||
struct OutPt;
|
||||
struct OutRec;
|
||||
struct Join;
|
||||
|
||||
typedef std::vector < OutRec* > PolyOutList;
|
||||
typedef std::vector < TEdge* > EdgeList;
|
||||
typedef std::vector < Join* > JoinList;
|
||||
typedef std::vector < IntersectNode* > IntersectList;
|
||||
|
||||
//------------------------------------------------------------------------------
|
||||
|
||||
//ClipperBase is the ancestor to the Clipper class. It should not be
|
||||
//instantiated directly. This class simply abstracts the conversion of sets of
|
||||
//polygon coordinates into edge objects that are stored in a LocalMinima list.
|
||||
class ClipperBase
|
||||
{
|
||||
public:
|
||||
ClipperBase();
|
||||
virtual ~ClipperBase();
|
||||
virtual bool AddPath(const Path &pg, PolyType PolyTyp, bool Closed);
|
||||
bool AddPaths(const Paths &ppg, PolyType PolyTyp, bool Closed);
|
||||
virtual void Clear();
|
||||
IntRect GetBounds();
|
||||
bool PreserveCollinear() {return m_PreserveCollinear;};
|
||||
void PreserveCollinear(bool value) {m_PreserveCollinear = value;};
|
||||
protected:
|
||||
void DisposeLocalMinimaList();
|
||||
TEdge* AddBoundsToLML(TEdge *e, bool IsClosed);
|
||||
virtual void Reset();
|
||||
TEdge* ProcessBound(TEdge* E, bool IsClockwise);
|
||||
void InsertScanbeam(const cInt Y);
|
||||
bool PopScanbeam(cInt &Y);
|
||||
bool LocalMinimaPending();
|
||||
bool PopLocalMinima(cInt Y, const LocalMinimum *&locMin);
|
||||
OutRec* CreateOutRec();
|
||||
void DisposeAllOutRecs();
|
||||
void DisposeOutRec(PolyOutList::size_type index);
|
||||
void SwapPositionsInAEL(TEdge *edge1, TEdge *edge2);
|
||||
void DeleteFromAEL(TEdge *e);
|
||||
void UpdateEdgeIntoAEL(TEdge *&e);
|
||||
|
||||
typedef std::vector<LocalMinimum> MinimaList;
|
||||
MinimaList::iterator m_CurrentLM;
|
||||
MinimaList m_MinimaList;
|
||||
|
||||
bool m_UseFullRange;
|
||||
EdgeList m_edges;
|
||||
bool m_PreserveCollinear;
|
||||
bool m_HasOpenPaths;
|
||||
PolyOutList m_PolyOuts;
|
||||
TEdge *m_ActiveEdges;
|
||||
|
||||
typedef std::priority_queue<cInt> ScanbeamList;
|
||||
ScanbeamList m_Scanbeam;
|
||||
};
|
||||
//------------------------------------------------------------------------------
|
||||
|
||||
class Clipper : public virtual ClipperBase
|
||||
{
|
||||
public:
|
||||
Clipper(int initOptions = 0);
|
||||
bool Execute(ClipType clipType,
|
||||
Paths &solution,
|
||||
PolyFillType fillType = pftEvenOdd);
|
||||
bool Execute(ClipType clipType,
|
||||
Paths &solution,
|
||||
PolyFillType subjFillType,
|
||||
PolyFillType clipFillType);
|
||||
bool Execute(ClipType clipType,
|
||||
PolyTree &polytree,
|
||||
PolyFillType fillType = pftEvenOdd);
|
||||
bool Execute(ClipType clipType,
|
||||
PolyTree &polytree,
|
||||
PolyFillType subjFillType,
|
||||
PolyFillType clipFillType);
|
||||
bool ReverseSolution() { return m_ReverseOutput; };
|
||||
void ReverseSolution(bool value) {m_ReverseOutput = value;};
|
||||
bool StrictlySimple() {return m_StrictSimple;};
|
||||
void StrictlySimple(bool value) {m_StrictSimple = value;};
|
||||
//set the callback function for z value filling on intersections (otherwise Z is 0)
|
||||
#ifdef use_xyz
|
||||
void ZFillFunction(ZFillCallback zFillFunc);
|
||||
#endif
|
||||
protected:
|
||||
virtual bool ExecuteInternal();
|
||||
private:
|
||||
JoinList m_Joins;
|
||||
JoinList m_GhostJoins;
|
||||
IntersectList m_IntersectList;
|
||||
ClipType m_ClipType;
|
||||
typedef std::list<cInt> MaximaList;
|
||||
MaximaList m_Maxima;
|
||||
TEdge *m_SortedEdges;
|
||||
bool m_ExecuteLocked;
|
||||
PolyFillType m_ClipFillType;
|
||||
PolyFillType m_SubjFillType;
|
||||
bool m_ReverseOutput;
|
||||
bool m_UsingPolyTree;
|
||||
bool m_StrictSimple;
|
||||
#ifdef use_xyz
|
||||
ZFillCallback m_ZFill; //custom callback
|
||||
#endif
|
||||
void SetWindingCount(TEdge& edge);
|
||||
bool IsEvenOddFillType(const TEdge& edge) const;
|
||||
bool IsEvenOddAltFillType(const TEdge& edge) const;
|
||||
void InsertLocalMinimaIntoAEL(const cInt botY);
|
||||
void InsertEdgeIntoAEL(TEdge *edge, TEdge* startEdge);
|
||||
void AddEdgeToSEL(TEdge *edge);
|
||||
bool PopEdgeFromSEL(TEdge *&edge);
|
||||
void CopyAELToSEL();
|
||||
void DeleteFromSEL(TEdge *e);
|
||||
void SwapPositionsInSEL(TEdge *edge1, TEdge *edge2);
|
||||
bool IsContributing(const TEdge& edge) const;
|
||||
bool IsTopHorz(const cInt XPos);
|
||||
void DoMaxima(TEdge *e);
|
||||
void ProcessHorizontals();
|
||||
void ProcessHorizontal(TEdge *horzEdge);
|
||||
void AddLocalMaxPoly(TEdge *e1, TEdge *e2, const IntPoint &pt);
|
||||
OutPt* AddLocalMinPoly(TEdge *e1, TEdge *e2, const IntPoint &pt);
|
||||
OutRec* GetOutRec(int idx);
|
||||
void AppendPolygon(TEdge *e1, TEdge *e2);
|
||||
void IntersectEdges(TEdge *e1, TEdge *e2, IntPoint &pt);
|
||||
OutPt* AddOutPt(TEdge *e, const IntPoint &pt);
|
||||
OutPt* GetLastOutPt(TEdge *e);
|
||||
bool ProcessIntersections(const cInt topY);
|
||||
void BuildIntersectList(const cInt topY);
|
||||
void ProcessIntersectList();
|
||||
void ProcessEdgesAtTopOfScanbeam(const cInt topY);
|
||||
void BuildResult(Paths& polys);
|
||||
void BuildResult2(PolyTree& polytree);
|
||||
void SetHoleState(TEdge *e, OutRec *outrec);
|
||||
void DisposeIntersectNodes();
|
||||
bool FixupIntersectionOrder();
|
||||
void FixupOutPolygon(OutRec &outrec);
|
||||
void FixupOutPolyline(OutRec &outrec);
|
||||
bool IsHole(TEdge *e);
|
||||
bool FindOwnerFromSplitRecs(OutRec &outRec, OutRec *&currOrfl);
|
||||
void FixHoleLinkage(OutRec &outrec);
|
||||
void AddJoin(OutPt *op1, OutPt *op2, const IntPoint offPt);
|
||||
void ClearJoins();
|
||||
void ClearGhostJoins();
|
||||
void AddGhostJoin(OutPt *op, const IntPoint offPt);
|
||||
bool JoinPoints(Join *j, OutRec* outRec1, OutRec* outRec2);
|
||||
void JoinCommonEdges();
|
||||
void DoSimplePolygons();
|
||||
void FixupFirstLefts1(OutRec* OldOutRec, OutRec* NewOutRec);
|
||||
void FixupFirstLefts2(OutRec* InnerOutRec, OutRec* OuterOutRec);
|
||||
void FixupFirstLefts3(OutRec* OldOutRec, OutRec* NewOutRec);
|
||||
#ifdef use_xyz
|
||||
void SetZ(IntPoint& pt, TEdge& e1, TEdge& e2);
|
||||
#endif
|
||||
};
|
||||
//------------------------------------------------------------------------------
|
||||
|
||||
class ClipperOffset
|
||||
{
|
||||
public:
|
||||
ClipperOffset(double miterLimit = 2.0, double roundPrecision = 0.25);
|
||||
~ClipperOffset();
|
||||
void AddPath(const Path& path, JoinType joinType, EndType endType);
|
||||
void AddPaths(const Paths& paths, JoinType joinType, EndType endType);
|
||||
void Execute(Paths& solution, double delta);
|
||||
void Execute(PolyTree& solution, double delta);
|
||||
void Clear();
|
||||
double MiterLimit;
|
||||
double ArcTolerance;
|
||||
private:
|
||||
Paths m_destPolys;
|
||||
Path m_srcPoly;
|
||||
Path m_destPoly;
|
||||
std::vector<DoublePoint> m_normals;
|
||||
double m_delta, m_sinA, m_sin, m_cos;
|
||||
double m_miterLim, m_StepsPerRad;
|
||||
IntPoint m_lowest;
|
||||
PolyNode m_polyNodes;
|
||||
|
||||
void FixOrientations();
|
||||
void DoOffset(double delta);
|
||||
void OffsetPoint(int j, int& k, JoinType jointype);
|
||||
void DoSquare(int j, int k);
|
||||
void DoMiter(int j, int k, double r);
|
||||
void DoRound(int j, int k);
|
||||
};
|
||||
//------------------------------------------------------------------------------
|
||||
|
||||
class clipperException : public std::exception
|
||||
{
|
||||
public:
|
||||
clipperException(const char* description): m_descr(description) {}
|
||||
virtual ~clipperException() throw() {}
|
||||
virtual const char* what() const throw() {return m_descr.c_str();}
|
||||
private:
|
||||
std::string m_descr;
|
||||
};
|
||||
//------------------------------------------------------------------------------
|
||||
|
||||
} //ClipperLib namespace
|
||||
|
||||
#endif //clipper_hpp
|
||||
|
||||
|
||||
@ -1,7 +1,7 @@
|
||||
#ifndef YOLOV5_COMMON_H_
|
||||
#define YOLOV5_COMMON_H_
|
||||
#ifndef DBNET_COMMON_H_
|
||||
#define DBNET_COMMON_H_
|
||||
|
||||
#include <iostream>s
|
||||
#include <iostream>
|
||||
#include <fstream>
|
||||
#include <map>
|
||||
#include <sstream>
|
||||
@ -120,27 +120,27 @@ ILayer* convBnLeaky(INetworkDefinition *network, std::map<std::string, Weights>&
|
||||
IActivationLayer* basicBlock(INetworkDefinition *network, std::map<std::string, Weights>& weightMap, ITensor& input, int inch, int outch, int stride, std::string lname) {
|
||||
Weights emptywts{ DataType::kFLOAT, nullptr, 0 };
|
||||
|
||||
IConvolutionLayer* conv1 = network->addConvolution(input, outch, DimsHW{ 3, 3 }, weightMap[lname + "conv1.weight"], emptywts);
|
||||
IConvolutionLayer* conv1 = network->addConvolutionNd(input, outch, DimsHW{ 3, 3 }, weightMap[lname + "conv1.weight"], emptywts);
|
||||
assert(conv1);
|
||||
conv1->setStride(DimsHW{ stride, stride });
|
||||
conv1->setPadding(DimsHW{ 1, 1 });
|
||||
conv1->setStrideNd(DimsHW{ stride, stride });
|
||||
conv1->setPaddingNd(DimsHW{ 1, 1 });
|
||||
|
||||
IScaleLayer* bn1 = addBatchNorm2d(network, weightMap, *conv1->getOutput(0), lname + "bn1", 1e-5);
|
||||
|
||||
IActivationLayer* relu1 = network->addActivation(*bn1->getOutput(0), ActivationType::kRELU);
|
||||
assert(relu1);
|
||||
|
||||
IConvolutionLayer* conv2 = network->addConvolution(*relu1->getOutput(0), outch, DimsHW{ 3, 3 }, weightMap[lname + "conv2.weight"], emptywts);
|
||||
IConvolutionLayer* conv2 = network->addConvolutionNd(*relu1->getOutput(0), outch, DimsHW{ 3, 3 }, weightMap[lname + "conv2.weight"], emptywts);
|
||||
assert(conv2);
|
||||
conv2->setPadding(DimsHW{ 1, 1 });
|
||||
conv2->setPaddingNd(DimsHW{ 1, 1 });
|
||||
|
||||
IScaleLayer* bn2 = addBatchNorm2d(network, weightMap, *conv2->getOutput(0), lname + "bn2", 1e-5);
|
||||
|
||||
IElementWiseLayer* ew1;
|
||||
if (inch != outch) {
|
||||
IConvolutionLayer* conv3 = network->addConvolution(input, outch, DimsHW{ 1, 1 }, weightMap[lname + "downsample.0.weight"], emptywts);
|
||||
IConvolutionLayer* conv3 = network->addConvolutionNd(input, outch, DimsHW{ 1, 1 }, weightMap[lname + "downsample.0.weight"], emptywts);
|
||||
assert(conv3);
|
||||
conv3->setStride(DimsHW{ stride, stride });
|
||||
conv3->setStrideNd(DimsHW{ stride, stride });
|
||||
IScaleLayer* bn3 = addBatchNorm2d(network, weightMap, *conv3->getOutput(0), lname + "downsample.1", 1e-5);
|
||||
ew1 = network->addElementWise(*bn3->getOutput(0), *bn2->getOutput(0), ElementWiseOperation::kSUM);
|
||||
}
|
||||
|
||||
227
dbnet/dbnet.cpp
227
dbnet/dbnet.cpp
@ -4,10 +4,15 @@
|
||||
#include "logging.h"
|
||||
#include "common.hpp"
|
||||
#include <math.h>
|
||||
#include "clipper.hpp"
|
||||
|
||||
#define USE_FP16 // comment out this if want to use FP32
|
||||
#define DEVICE 0 // GPU id
|
||||
#define EXPANDRATIO 1.4
|
||||
#define EXPANDRATIO 1.5
|
||||
#define BOX_MINI_SIZE 5
|
||||
#define SCORE_THRESHOLD 0.3
|
||||
#define BOX_THRESHOLD 0.7
|
||||
|
||||
static const int SHORT_INPUT = 640;
|
||||
static const int MAX_INPUT_SIZE = 1440; // 32x
|
||||
static const int MIN_INPUT_SIZE = 608;
|
||||
@ -17,12 +22,37 @@ const char* INPUT_BLOB_NAME = "data";
|
||||
const char* OUTPUT_BLOB_NAME = "out";
|
||||
static Logger gLogger;
|
||||
|
||||
cv::RotatedRect expandBox(const cv::RotatedRect& inBox, float ratio = 1.0) {
|
||||
cv::Size size = inBox.size;
|
||||
int neww = size.width * ratio;
|
||||
int newh = size.height *ratio;
|
||||
return cv::RotatedRect(inBox.center, cv::Size(neww, newh), inBox.angle);
|
||||
cv::RotatedRect expandBox(cv::Point2f temp[], float ratio)
|
||||
{
|
||||
ClipperLib::Path path = {
|
||||
{ClipperLib::cInt(temp[0].x), ClipperLib::cInt(temp[0].y)},
|
||||
{ClipperLib::cInt(temp[1].x), ClipperLib::cInt(temp[1].y)},
|
||||
{ClipperLib::cInt(temp[2].x), ClipperLib::cInt(temp[2].y)},
|
||||
{ClipperLib::cInt(temp[3].x), ClipperLib::cInt(temp[3].y)}};
|
||||
double area = ClipperLib::Area(path);
|
||||
double distance;
|
||||
double length = 0.0;
|
||||
for (int i = 0; i < 4; i++) {
|
||||
length = length + sqrtf(powf((temp[i].x - temp[(i + 1) % 4].x), 2) +
|
||||
powf((temp[i].y - temp[(i + 1) % 4].y), 2));
|
||||
}
|
||||
|
||||
distance = area * ratio / length;
|
||||
|
||||
ClipperLib::ClipperOffset offset;
|
||||
offset.AddPath(path, ClipperLib::JoinType::jtRound,
|
||||
ClipperLib::EndType::etClosedPolygon);
|
||||
ClipperLib::Paths paths;
|
||||
offset.Execute(paths, distance);
|
||||
|
||||
std::vector<cv::Point> contour;
|
||||
for (int i = 0; i < paths[0].size(); i++) {
|
||||
contour.emplace_back(paths[0][i].X, paths[0][i].Y);
|
||||
}
|
||||
offset.Clear();
|
||||
return cv::minAreaRect(contour);
|
||||
}
|
||||
|
||||
float paddimg(cv::Mat& In_Out_img, int shortsize = 960) {
|
||||
int w = In_Out_img.cols;
|
||||
int h = In_Out_img.rows;
|
||||
@ -48,6 +78,7 @@ float paddimg(cv::Mat& In_Out_img, int shortsize = 960) {
|
||||
cv::resize(In_Out_img, In_Out_img, cv::Size(w, h));
|
||||
return scale;
|
||||
}
|
||||
|
||||
// Creat the engine using only the API and not any parser.
|
||||
ICudaEngine* createEngine(unsigned int maxBatchSize, IBuilder* builder, IBuilderConfig* config, DataType dt) {
|
||||
const auto explicitBatch = 1U << static_cast<uint32_t>(NetworkDefinitionCreationFlag::kEXPLICIT_BATCH);
|
||||
@ -56,23 +87,23 @@ ICudaEngine* createEngine(unsigned int maxBatchSize, IBuilder* builder, IBuilder
|
||||
ITensor* data = network->addInput(INPUT_BLOB_NAME, dt, Dims4{ 1, 3, -1, -1 });
|
||||
assert(data);
|
||||
|
||||
std::map<std::string, Weights> weightMap = loadWeights("E:\\LearningCodes\\DBNET\\DBNet.pytorch\\tools\\DBNet.wts");
|
||||
std::map<std::string, Weights> weightMap = loadWeights("./DBNet.wts");
|
||||
Weights emptywts{ DataType::kFLOAT, nullptr, 0 };
|
||||
|
||||
/* ------ Resnet18 backbone------ */
|
||||
// Add convolution layer with 6 outputs and a 5x5 filter.
|
||||
IConvolutionLayer* conv1 = network->addConvolution(*data, 64, DimsHW{ 7, 7 }, weightMap["backbone.conv1.weight"], emptywts);
|
||||
IConvolutionLayer* conv1 = network->addConvolutionNd(*data, 64, DimsHW{ 7, 7 }, weightMap["backbone.conv1.weight"], emptywts);
|
||||
assert(conv1);
|
||||
conv1->setStride(DimsHW{ 2, 2 });
|
||||
conv1->setPadding(DimsHW{ 3, 3 });
|
||||
conv1->setStrideNd(DimsHW{ 2, 2 });
|
||||
conv1->setPaddingNd(DimsHW{ 3, 3 });
|
||||
|
||||
IScaleLayer* bn1 = addBatchNorm2d(network, weightMap, *conv1->getOutput(0), "backbone.bn1", 1e-5);
|
||||
IActivationLayer* relu1 = network->addActivation(*bn1->getOutput(0), ActivationType::kRELU);
|
||||
assert(relu1);
|
||||
IPoolingLayer* pool1 = network->addPooling(*relu1->getOutput(0), PoolingType::kMAX, DimsHW{ 3, 3 });
|
||||
IPoolingLayer* pool1 = network->addPoolingNd(*relu1->getOutput(0), PoolingType::kMAX, DimsHW{ 3, 3 });
|
||||
assert(pool1);
|
||||
pool1->setStride(DimsHW{ 2, 2 });
|
||||
pool1->setPadding(DimsHW{ 1, 1 });
|
||||
pool1->setStrideNd(DimsHW{ 2, 2 });
|
||||
pool1->setPaddingNd(DimsHW{ 1, 1 });
|
||||
|
||||
IActivationLayer* relu2 = basicBlock(network, weightMap, *pool1->getOutput(0), 64, 64, 1, "backbone.layer1.0.");
|
||||
IActivationLayer* relu3 = basicBlock(network, weightMap, *relu2->getOutput(0), 64, 64, 1, "backbone.layer1.1."); // x2
|
||||
@ -132,7 +163,7 @@ ICudaEngine* createEngine(unsigned int maxBatchSize, IBuilder* builder, IBuilder
|
||||
}
|
||||
Weights deconvwts5{ DataType::kFLOAT, deval2, 64 * 8 * 8 };
|
||||
IDeconvolutionLayer* p4_up_p2 = network->addDeconvolutionNd(*p4->getOutput(0), 64, DimsHW{ 8, 8 }, deconvwts5, emptywts);
|
||||
p4_up_p2->setPadding(DimsHW{ 2, 2 });
|
||||
p4_up_p2->setPaddingNd(DimsHW{ 2, 2 });
|
||||
p4_up_p2->setStrideNd(DimsHW{ 4, 4 });
|
||||
p4_up_p2->setNbGroups(64);
|
||||
weightMap["deconv2"] = deconvwts5;
|
||||
@ -162,10 +193,10 @@ ICudaEngine* createEngine(unsigned int maxBatchSize, IBuilder* builder, IBuilder
|
||||
binarizeup2->setStrideNd(DimsHW{ 2, 2 });
|
||||
binarizeup2->setNbGroups(64);
|
||||
|
||||
IConvolutionLayer* binarize3 = network->addConvolution(*binarizeup2->getOutput(0), 1, DimsHW{ 3, 3 }, weightMap["head.binarize.7.weight"], weightMap["head.binarize.7.bias"]);
|
||||
IConvolutionLayer* binarize3 = network->addConvolutionNd(*binarizeup2->getOutput(0), 1, DimsHW{ 3, 3 }, weightMap["head.binarize.7.weight"], weightMap["head.binarize.7.bias"]);
|
||||
assert(binarize3);
|
||||
binarize3->setStride(DimsHW{ 1, 1 });
|
||||
binarize3->setPadding(DimsHW{ 1, 1 });
|
||||
binarize3->setStrideNd(DimsHW{ 1, 1 });
|
||||
binarize3->setPaddingNd(DimsHW{ 1, 1 });
|
||||
IActivationLayer* binarize4 = network->addActivation(*binarize3->getOutput(0), ActivationType::kSIGMOID);
|
||||
assert(binarize4);
|
||||
|
||||
@ -175,10 +206,10 @@ ICudaEngine* createEngine(unsigned int maxBatchSize, IBuilder* builder, IBuilder
|
||||
IDeconvolutionLayer* threshup = network->addDeconvolutionNd(*thresh1->getOutput(0), 64, DimsHW{ 2, 2 }, deconvwts9, emptywts);
|
||||
threshup->setStrideNd(DimsHW{ 2, 2 });
|
||||
threshup->setNbGroups(64);
|
||||
IConvolutionLayer* thresh2 = network->addConvolution(*threshup->getOutput(0), 64, DimsHW{ 3, 3 }, weightMap["head.thresh.3.1.weight"], weightMap["head.thresh.3.1.bias"]);
|
||||
IConvolutionLayer* thresh2 = network->addConvolutionNd(*threshup->getOutput(0), 64, DimsHW{ 3, 3 }, weightMap["head.thresh.3.1.weight"], weightMap["head.thresh.3.1.bias"]);
|
||||
assert(thresh2);
|
||||
thresh2->setStride(DimsHW{ 1, 1 });
|
||||
thresh2->setPadding(DimsHW{ 1, 1 });
|
||||
thresh2->setStrideNd(DimsHW{ 1, 1 });
|
||||
thresh2->setPaddingNd(DimsHW{ 1, 1 });
|
||||
|
||||
IScaleLayer* threshbn1 = addBatchNorm2d(network, weightMap, *thresh2->getOutput(0), "head.thresh.4", 1e-5);
|
||||
IActivationLayer* threshrelu1 = network->addActivation(*threshbn1->getOutput(0), ActivationType::kRELU);
|
||||
@ -188,10 +219,10 @@ ICudaEngine* createEngine(unsigned int maxBatchSize, IBuilder* builder, IBuilder
|
||||
IDeconvolutionLayer* threshup2 = network->addDeconvolutionNd(*threshrelu1->getOutput(0), 64, DimsHW{ 2, 2 }, deconvwts10, emptywts);
|
||||
threshup2->setStrideNd(DimsHW{ 2, 2 });
|
||||
threshup2->setNbGroups(64);
|
||||
IConvolutionLayer* thresh3 = network->addConvolution(*threshup2->getOutput(0), 1, DimsHW{ 3, 3 }, weightMap["head.thresh.6.1.weight"], weightMap["head.thresh.6.1.bias"]);
|
||||
IConvolutionLayer* thresh3 = network->addConvolutionNd(*threshup2->getOutput(0), 1, DimsHW{ 3, 3 }, weightMap["head.thresh.6.1.weight"], weightMap["head.thresh.6.1.bias"]);
|
||||
assert(thresh3);
|
||||
thresh3->setStride(DimsHW{ 1, 1 });
|
||||
thresh3->setPadding(DimsHW{ 1, 1 });
|
||||
thresh3->setStrideNd(DimsHW{ 1, 1 });
|
||||
thresh3->setPaddingNd(DimsHW{ 1, 1 });
|
||||
IActivationLayer* thresh4 = network->addActivation(*thresh3->getOutput(0), ActivationType::kSIGMOID);
|
||||
assert(thresh4);
|
||||
|
||||
@ -281,6 +312,91 @@ void doInference(IExecutionContext& context, float* input, float* output, int h_
|
||||
CHECK(cudaFree(buffers[outputIndex]));
|
||||
}
|
||||
|
||||
bool get_mini_boxes(cv::RotatedRect& rotated_rect, cv::Point2f rect[],
|
||||
int min_size)
|
||||
{
|
||||
|
||||
cv::Point2f temp_rect[4];
|
||||
rotated_rect.points(temp_rect);
|
||||
for (int i = 0; i < 4; i++) {
|
||||
for (int j = i + 1; j < 4; j++) {
|
||||
if (temp_rect[i].x > temp_rect[j].x) {
|
||||
cv::Point2f temp;
|
||||
temp = temp_rect[i];
|
||||
temp_rect[i] = temp_rect[j];
|
||||
temp_rect[j] = temp;
|
||||
}
|
||||
}
|
||||
}
|
||||
int index0 = 0;
|
||||
int index1 = 1;
|
||||
int index2 = 2;
|
||||
int index3 = 3;
|
||||
if (temp_rect[1].y > temp_rect[0].y) {
|
||||
index0 = 0;
|
||||
index3 = 1;
|
||||
} else {
|
||||
index0 = 1;
|
||||
index3 = 0;
|
||||
}
|
||||
if (temp_rect[3].y > temp_rect[2].y) {
|
||||
index1 = 2;
|
||||
index2 = 3;
|
||||
} else {
|
||||
index1 = 3;
|
||||
index2 = 2;
|
||||
}
|
||||
|
||||
rect[0] = temp_rect[index0]; // Left top coordinate
|
||||
rect[1] = temp_rect[index1]; // Left bottom coordinate
|
||||
rect[2] = temp_rect[index2]; // Right bottom coordinate
|
||||
rect[3] = temp_rect[index3]; // Right top coordinate
|
||||
|
||||
if (rotated_rect.size.width < min_size ||
|
||||
rotated_rect.size.height < min_size) {
|
||||
return false;
|
||||
} else {
|
||||
return true;
|
||||
}
|
||||
}
|
||||
|
||||
float get_box_score(float* map, cv::Point2f rect[], int width, int height,
|
||||
float threshold)
|
||||
{
|
||||
|
||||
int xmin = width - 1;
|
||||
int ymin = height - 1;
|
||||
int xmax = 0;
|
||||
int ymax = 0;
|
||||
|
||||
for (int j = 0; j < 4; j++) {
|
||||
if (rect[j].x < xmin) {
|
||||
xmin = rect[j].x;
|
||||
}
|
||||
if (rect[j].y < ymin) {
|
||||
ymin = rect[j].y;
|
||||
}
|
||||
if (rect[j].x > xmax) {
|
||||
xmax = rect[j].x;
|
||||
}
|
||||
if (rect[j].y > ymax) {
|
||||
ymax = rect[j].y;
|
||||
}
|
||||
}
|
||||
float sum = 0;
|
||||
int num = 0;
|
||||
for (int i = ymin; i <= ymax; i++) {
|
||||
for (int j = xmin; j <= xmax; j++) {
|
||||
if (map[i * width + j] > threshold) {
|
||||
sum = sum + map[i * width + j];
|
||||
num++;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
return sum / num;
|
||||
}
|
||||
|
||||
int main(int argc, char** argv) {
|
||||
cudaSetDevice(DEVICE);
|
||||
// create a model using the API directly and serialize it to a stream
|
||||
@ -334,19 +450,24 @@ int main(int argc, char** argv) {
|
||||
return -1;
|
||||
}
|
||||
|
||||
// icdar2015.yaml Hyperparameter
|
||||
std::vector<float> mean_value{ 0.406, 0.456, 0.485 }; // BGR
|
||||
std::vector<float> std_value{ 0.225, 0.224, 0.229 };
|
||||
|
||||
int fcount = 0;
|
||||
|
||||
for (auto f : file_names) {
|
||||
fcount++;
|
||||
std::cout << fcount << " " << f << std::endl;
|
||||
cv::Mat pr_img = cv::imread(std::string(argv[2]) + "/" + f);
|
||||
cv::Mat src_img = pr_img.clone();
|
||||
if (pr_img.empty()) continue;
|
||||
float scale = paddimg(pr_img, SHORT_INPUT);
|
||||
float scale = paddimg(pr_img, SHORT_INPUT); // resize the image
|
||||
std::cout << "letterbox shape: " << pr_img.cols << ", " << pr_img.rows << std::endl;
|
||||
if (pr_img.cols < MIN_INPUT_SIZE || pr_img.rows < MIN_INPUT_SIZE) continue;
|
||||
float* data = new float[3 * pr_img.rows * pr_img.cols];
|
||||
|
||||
auto start = std::chrono::system_clock::now();
|
||||
int i = 0;
|
||||
for (int row = 0; row < pr_img.rows; ++row) {
|
||||
uchar* uc_pixel = pr_img.data + row * pr_img.step;
|
||||
@ -358,15 +479,17 @@ int main(int argc, char** argv) {
|
||||
++i;
|
||||
}
|
||||
}
|
||||
auto end = std::chrono::system_clock::now();
|
||||
std::cout << "pre time:"<< std::chrono::duration_cast<std::chrono::milliseconds>(end - start).count() << "ms" << std::endl;
|
||||
|
||||
float* prob = new float[pr_img.rows *pr_img.cols * 2];
|
||||
// Run inference
|
||||
auto start = std::chrono::system_clock::now();
|
||||
start = std::chrono::system_clock::now();
|
||||
doInference(*context, data, prob, pr_img.rows, pr_img.cols);
|
||||
auto end = std::chrono::system_clock::now();
|
||||
std::cout << std::chrono::duration_cast<std::chrono::milliseconds>(end - start).count() << "ms" << std::endl;
|
||||
end = std::chrono::system_clock::now();
|
||||
std::cout << "detect time:"<< std::chrono::duration_cast<std::chrono::milliseconds>(end - start).count() << "ms" << std::endl;
|
||||
|
||||
// prob 为 2* 640*640 拿出第一个
|
||||
// prob shape is 2*640*640, get the first one
|
||||
cv::Mat map = cv::Mat::zeros(cv::Size(pr_img.cols, pr_img.rows), CV_8UC1);
|
||||
for (int h = 0; h < pr_img.rows; ++h) {
|
||||
uchar *ptr = map.ptr(h);
|
||||
@ -374,7 +497,8 @@ int main(int argc, char** argv) {
|
||||
ptr[w] = (prob[h * pr_img.cols + w] > 0.3) ? 255 : 0;
|
||||
}
|
||||
}
|
||||
// 提取最小外接矩形
|
||||
|
||||
// Extracting minimum circumscribed rectangle
|
||||
std::vector<std::vector<cv::Point>> contours;
|
||||
std::vector<cv::Vec4i> hierarcy;
|
||||
cv::findContours(map, contours, hierarcy, CV_RETR_LIST, CV_CHAIN_APPROX_SIMPLE);
|
||||
@ -382,24 +506,43 @@ int main(int argc, char** argv) {
|
||||
std::vector<cv::Rect> boundRect(contours.size());
|
||||
std::vector<cv::RotatedRect> box(contours.size());
|
||||
cv::Point2f rect[4];
|
||||
cv::Point2f order_rect[4];
|
||||
|
||||
for (int i = 0; i < contours.size(); i++) {
|
||||
box[i] = cv::minAreaRect(cv::Mat(contours[i]));
|
||||
//boundRect[i] = cv::boundingRect(cv::Mat(contours[i]));
|
||||
//绘制外接矩形和 最小外接矩形（for循环）
|
||||
//cv::rectangle(img, cv::Point(boundRect[i].x, boundRect[i].y), cv::Point(boundRect[i].x + boundRect[i].width, boundRect[i].y + boundRect[i].height), cv::Scalar(0, 255, 0), 2, 8);
|
||||
cv::RotatedRect expandbox = expandBox(box[i], EXPANDRATIO);
|
||||
expandbox.points(rect);//把最小外接矩形四个端点复制给rect数组
|
||||
for (int j = 0; j < 4; j++) {
|
||||
cv::Point2f p1, p2;
|
||||
p1.x = round(rect[j].x / pr_img.cols * src_img.cols);
|
||||
p1.y = round(rect[j].y / pr_img.rows * src_img.rows);
|
||||
p2.x = round(rect[(j + 1) % 4].x / pr_img.cols * src_img.cols);
|
||||
p2.y = round(rect[(j + 1) % 4].y / pr_img.rows * src_img.rows);
|
||||
cv::line(src_img, p1, p2, cv::Scalar(0, 0, 255), 2, 8);
|
||||
cv::RotatedRect rotated_rect = cv::minAreaRect(cv::Mat(contours[i]));
|
||||
if (!get_mini_boxes(rotated_rect, rect, BOX_MINI_SIZE)) {
|
||||
std::cout << "box too small" << std::endl;
|
||||
continue;
|
||||
}
|
||||
|
||||
// drop low score boxes
|
||||
float score = get_box_score(prob, rect, pr_img.cols, pr_img.rows,
|
||||
SCORE_THRESHOLD);
|
||||
if (score < BOX_THRESHOLD) {
|
||||
std::cout << "score too low = " << score << ", threshold = " << BOX_THRESHOLD << std::endl;
|
||||
continue;
|
||||
}
|
||||
|
||||
// Scaling the predict boxes depend on EXPANDRATIO
|
||||
cv::RotatedRect expandbox = expandBox(rect, EXPANDRATIO);
|
||||
expandbox.points(rect);
|
||||
if (!get_mini_boxes(expandbox, rect, BOX_MINI_SIZE + 2)) {
|
||||
continue;
|
||||
}
|
||||
|
||||
// Restore the coordinates to the original image
|
||||
for (int k = 0; k < 4; k++) {
|
||||
order_rect[k] = rect[k];
|
||||
order_rect[k].x = int(order_rect[k].x / pr_img.cols * src_img.cols);
|
||||
order_rect[k].y = int(order_rect[k].y / pr_img.rows * src_img.rows);
|
||||
}
|
||||
|
||||
cv::rectangle(src_img, cv::Point(order_rect[0].x,order_rect[0].y), cv::Point(order_rect[2].x,order_rect[2].y), cv::Scalar(0, 0, 255), 2, 8);
|
||||
//std::cout << "After LT = " << order_rect[0] << ", After RD = " << order_rect[2] << std::endl;
|
||||
}
|
||||
|
||||
cv::imwrite("_" + f, src_img);
|
||||
std::cout << "write image done." << std::endl;
|
||||
//cv::waitKey(0);
|
||||
|
||||
delete prob;
|
||||
|
||||
Loading…
Reference in New Issue
Block a user