source: proiecte/pmake3d/make3d_original/Make3dSingleImageStanford_version0.1/LearningCode/Learning/OldBatchVersion/LearnDisCts.m @ 37

Last change on this file since 37 was 37, checked in by (none), 14 years ago

Added original make3d
File size: 15.8 KB
Line 
1% *  This code was used in the following articles:
2% *  [1] Learning 3-D Scene Structure from a Single Still Image,
3% *      Ashutosh Saxena, Min Sun, Andrew Y. Ng,
4% *      In ICCV workshop on 3D Representation for Recognition (3dRR-07), 2007.
5% *      (best paper)
6% *  [2] 3-D Reconstruction from Sparse Views using Monocular Vision,
7% *      Ashutosh Saxena, Min Sun, Andrew Y. Ng,
8% *      In ICCV workshop on Virtual Representations and Modeling
9% *      of Large-scale environments (VRML), 2007.
10% *  [3] 3-D Depth Reconstruction from a Single Still Image,
11% *      Ashutosh Saxena, Sung H. Chung, Andrew Y. Ng.
12% *      International Journal of Computer Vision (IJCV), Aug 2007.
13% *  [6] Learning Depth from Single Monocular Images,
14% *      Ashutosh Saxena, Sung H. Chung, Andrew Y. Ng.
15% *      In Neural Information Processing Systems (NIPS) 18, 2005.
16% *
17% *  These articles are available at:
18% *  http://make3d.stanford.edu/publications
19% *
20% *  We request that you cite the papers [1], [3] and [6] in any of
21% *  your reports that uses this code.
22% *  Further, if you use the code in image3dstiching/ (multiple image version),
23% *  then please cite [2].
24% * 
25% *  If you use the code in third_party/, then PLEASE CITE and follow the
26% *  LICENSE OF THE CORRESPONDING THIRD PARTY CODE.
27% *
28% *  Finally, this code is for non-commercial use only.  For further
29% *  information and to obtain a copy of the license, see
30% *
31% *  http://make3d.stanford.edu/publications/code
32% *
33% *  Also, the software distributed under the License is distributed on an
34% * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either
35% *  express or implied.   See the License for the specific language governing
36% *  permissions and limitations under the License.
37% *
38% */
39function LearnDisCts(LearnAlg,HistFeaType,HistFeaDate,AbsFeaType,AbsFeaDate,WeiBatchNumber,logScale,SkyExclude,LearnNear);
40%WeiBatchNumber,logScale,SkyExclude)
41% % This function learned the distance
42
43global GeneralDataFolder ScratchDataFolder LocalFolder ClusterExecutionDirectory...
44    ImgFolder VertYNuPatch VertYNuDepth HoriXNuPatch HoriXNuDepth a_default b_default Ox_default Oy_default...
45    Horizon_default filename batchSize NuRow_default WeiBatchSize;
46
47statusFilename = [ClusterExecutionDirectory '/matlabExecutionStatus_ratio.txt'];
48% parameters setting
49NuPics = size(filename,2);
50NuBatch = ceil(NuPics/batchSize);
51NuRow = NuRow_default;
52% =================== can be modified later==============
53SpiatalDisCTSThresh = log(5);
54% =======================================================
55%Horizon = Horizon_default;
56%skyBottom = floor(NuRow/2);
57batchRow = 1:WeiBatchSize:NuRow;
58
59% constructing features for each batch of rows from batch featuresa
60%load([ScratchDataFolder '/data/FeatureSuperpixel.mat']); % load the feature relate to position and shape of superpixel
61% load estimated sky
62load([ScratchDataFolder '/data/MaskGSky.mat']); % maskg is the estimated ground maskSky is the estimated sky
63% load data
64load([ScratchDataFolder '/data/LowResImgIndexSuperpixelSep.mat']); % LowResImgIndexSuperpixelSep
65load([ScratchDataFolder '/data/DiffLowResImgIndexSuperpixelSep.mat']); % DiffLowResImgIndexSuperpixelSep(medi$large)
66%load([ScratchDataFolder '/data/TextLowResImgIndexSuperpixelSep.mat']); % TextLowResImgIndexSuperpixelSep using Textrea
67
68l = 1;
69for i = batchRow(WeiBatchNumber):min(batchRow(WeiBatchNumber)+WeiBatchSize-1,NuRow)
70    %i = NuRow
71    l
72    FeaVectorH = [];
73    FeaVectorV = [];
74%    FeaWeiH = [];
75%    FeaWeiV = [];
76    DepthDisCtsVectorH = [];
77    DepthDisCtsVectorV = [];
78    fid = fopen(statusFilename, 'w+');
79    fprintf(fid, 'Currently on row number %i\n', i);
80    fclose(fid);        %file opening and closing has to be inside the loop, otherwise the file will not appear over afs
81    for j = 1:NuBatch
82        tic
83        load([ScratchDataFolder '/data/feature_Abs_' AbsFeaType int2str(j) '_' AbsFeaDate '.mat']); % 'f'
84 %       load([ScratchDataFolder '/data/feature_Hist_' HistFeaType int2str(j) '_' HistFeaDate '.mat']); % 'f'
85        %toc
86        %for k = trainIndex{j}
87        for k = 1:size(f,2)%batchSize
88           
89            %==================
90            % load picsinfo just for the horizontal value
91            PicsinfoName = strrep(filename{(j-1)*batchSize+k},'img','picsinfo');
92            temp = dir([GeneralDataFolder '/PicsInfo/' PicsinfoName '.mat']);
93            if size(temp,1) == 0
94                a = a_default;
95                b = b_default;
96                Ox = Ox_default;
97                Oy = Oy_default;
98                Horizon = Horizon_default;
99            else   
100                load([GeneralDataFolder '/PicsInfo/' PicsinfoName '.mat']);
101            end
102
103            % load the depths
104            depthfile = strrep(filename{(j-1)*batchSize+k},'img','depth_sph_corr');
105            load([ScratchDataFolder '/Gridlaserdata/' depthfile '.mat']);
106            %==================
107            %tic;
108            % generate the range of the row for the same thi (weight value)
109            RowskyBottom = ceil(NuRow/2);
110            PatchSkyBottom = ceil(VertYNuDepth*(1-Horizon));
111            if i <= RowskyBottom
112               PatchRowRatio = PatchSkyBottom/RowskyBottom;
113               RowTop = ceil((i-1)*PatchRowRatio+1);
114               RowBottom = ceil(i*PatchRowRatio);
115            else
116               PatchRowRatio = (VertYNuDepth-PatchSkyBottom)/(NuRow-RowskyBottom);
117               RowTop = ceil((i-RowskyBottom-1)*PatchRowRatio+1)+PatchSkyBottom;
118               RowBottom = ceil((i-RowskyBottom)*PatchRowRatio)+PatchSkyBottom;
119            end
120            ColumnLeft = 1;
121            ColumnRight = HoriXNuDepth;
122           
123            % generate the DisCts
124            [DisCtsH DisCtsV] = DisCtsDetect(Position3DGrid(:,:,4));
125            % pick out the 1st end-1 end column and row
126            rangeYV = RowTop:RowBottom;
127            rangeYH = ColumnLeft:ColumnRight;
128            %rangeYV = setdiff(rangeYV,[1 VertYNuDepth-1 VertYNuDepth]);
129            rangeXV = RowTop:RowBottom;
130            rangeXH = ColumnLeft:ColumnRight;
131            %rangeXH = setdiff(rangeXH,[1 HoriXNuDepth-1 HoriXNuDepth]);
132            temp = DisCtsV(rangeYV,rangeYH);
133            DepthDisCtsVectorV = [DepthDisCtsVectorV; temp(:)];
134            temp = DisCtsH(rangeXV,rangeXH);
135            DepthDisCtsVectorH = [DepthDisCtsVectorH; temp(:)];
136
137            % pick out the features in the right rows   
138            newFeaH = genFeaVector(f{k},[],...
139                     rangeXV,rangeXH,(j-1)*batchSize+k,LearnNear,[1]);
140            newFeaH = abs(newFeaH - genFeaVector(f{k},[],...
141                     rangeXV,(rangeXH+1),(j-1)*batchSize+k,LearnNear,[1]));
142            newFeaV = genFeaVector(f{k},[],...
143                     rangeYV,rangeYH,(j-1)*batchSize+k,LearnNear,[1]);
144            newFeaV = abs(newFeaV - genFeaVector(f{k},[],...
145                     (rangeYV+1),rangeYH,(j-1)*batchSize+k,LearnNear,[1]));
146% ==================================
147        % setting the estimated Ground Verticl Sky segmentation (generated from CMU's code)
148        maskEstGVS = 2*ones(VertYNuDepth,HoriXNuDepth);
149        maskEstGVS(maskg{i}) = 1;
150        maskEstGVS(maskSky{i}) = 3;
151
152            % generate segmentation features
153        NuSupType = size(LowResImgIndexSuperpixelSep,2)+size(DiffLowResImgIndexSuperpixelSep,2);
154                   %+size(TextLowResImgIndexSuperpixelSep,2)*size(TextLowResImgIndexSuperpixelSep,3);
155
156        for j = 1:NuSupType  % total 21 seg : 3 RGB 6*3= 18 texture filters
157
158            % pick the specific segmentation
159            if j==1
160                sup = LowResImgIndexSuperpixelSep{i,1};
161            elseif j<=3
162                sup = DiffLowResImgIndexSuperpixelSep{i,j-1};
163            else
164                Subm = mod((j-3),6);
165                if Subm==0
166                   Subm=6;
167                end
168                sup = TextLowResImgIndexSuperpixelSep{i,Subm,ceil((j-3)/6)};
169            end
170
171            % extend the estimated maskGVS to the new segmentation
172            NewSupInd = (unique(sup))';
173            NewEstGSup = zeros(VertYNuDepth,HoriXNuDepth);
174            NewEstVSup = zeros(VertYNuDepth,HoriXNuDepth);
175            NewEstSSup = zeros(VertYNuDepth,HoriXNuDepth);
176            for m = NewSupInd
177                mask = sup == m;
178                if any(maskEstGVS(mask)==1) && any(maskEstGVS(mask)==3)
179                    GVSInd = analysesupinpatch(maskEstGVS(mask));
180                elseif any(maskEstGVS(mask)==1)
181                    GVSInd =1;
182                elseif any(maskEstGVS(mask)==3)
183                    GVSInd =3;
184                else
185                    GVSInd =2;
186                end
187                %GVSInd = analysesupinpatch(maskEstGVS(mask));
188                if GVSInd == 1
189                   NewEstGSup(mask) = m;
190                   NewEstVSup(mask) = 0;
191                   NewEstSSup(mask) = 0;
192                elseif GVSInd == 2
193                   NewEstVSup(mask) = m;
194                   NewEstGSup(mask) = -1;
195                   NewEstSSup(mask) = -1;
196                else
197                   NewEstSSup(mask) = m;
198                   NewEstGSup(mask) = -2;
199                   NewEstVSup(mask) = -2;
200                end
201            end
202            %if j == 2
203            %   SpreadFactor = gen_SFactor(LearnedDepth,sup,Rz);
204            %end
205            %clear LowResImgIndexSuperpixelSep;
206
207            % 2nd order smooth
208%            [SecXG(j,:) SecYG(j,:)]= gen_2ndSmooth(NewEstGSup);
209%            [SecXV(j,:) SecYV(j,:)]= gen_2ndSmooth(NewEstVSup);
210%            [SecXS(j,:) SecYS(j,:)]= gen_2ndSmooth(NewEstSSup);
211
212            % 1st order smooth
213            [FirstYG(j,:) FirstXG(j,:)] = gen_1stSmooth(NewEstGSup);
214            [FirstYV(j,:) FirstXV(j,:)] = gen_1stSmooth(NewEstVSup);
215            [FirstYS(j,:) FirstXS(j,:)] = gen_1stSmooth(NewEstSSup);
216             FirstY = FirstYG + FirstYV + FirstYS;
217             FirstX = FirstXG + FirstXV + FirstXS;
218             
219                %[GPy{j} ] = gen_GravityP_vertical(maskV);
220            %[PlanePriorX PlanePriorY]= gen_PlanePrior(LowResImgIndexSuperpixelSep{i,1});
221        end
222           
223            %size(newFeaH)
224            % add on the Hist features
225            %SizeHisFea = size(RFVector{k},3);
226            %HistFea = genHisFeaVector(reshape(RFVector{k},[],SizeHisFea),...
227            %          RowTop,RowBottom,ColumnLeft,ColumnRight,(j-1)*batchSize+k);
228            %rank(HistFea)
229            %newFeaH = [newFeaH; HistFea];
230            %newFeaV = [newFeaV; HistFea];
231           
232            % finally storage all the data
233            [Ix Iy] = meshgrid(rangeXH,rangeXV);
234            Index = sub2ind([VertYNuDepth, HoriXNuDepth], Iy(:), Ix(:));
235            FeaVectorH =[ FeaVectorH [newFeaH; FirstX(:,Index) ]];
236            FeaVectorV =[ FeaVectorV [newFeaV; FirstY(:,Index) ]];
237        end
238            % get rid of the sky region
239        clear f newFeaH newFeaV DisCtsH DisCtsV Position3DGrid;
240        toc
241    end
242
243    % learning part
244       targetH{l} = DepthDisCtsVectorH(:);
245       targetV{l} = DepthDisCtsVectorV(:);
246    size(FeaVectorH)
247    TsizeH = size(targetH{l},1)
248    FsizeH = size(FeaVectorH,1)+1
249    TsizeV = size(targetV{l},1)
250    FsizeV = size(FeaVectorV,1)+1
251    % calculate the weight to even the bias of much more spatial smooth data
252    wetH = ones(TsizeH,1);
253    wetH(targetH{l}==1) = sum(targetH{l}==0)./sum(targetH{l}==1);
254    wetV = ones(TsizeV,1);
255    wetH(targetV{l}==1) = sum(targetV{l}==0)./sum(targetV{l}==1);
256   
257   
258
259    % start learning ========================================
260    if strcmp(LearnAlg,'Logit')
261       size([targetH{l} ones(TsizeH,1)])
262       [thiH{l}] = glmfit(FeaVectorH', [targetH{l} ones(TsizeH,1)],'binomial', 'link', 'logit','weights',wetH);
263       X_baseH = ones(TsizeH,1);
264       [thi_baseH{l}] = glmfit(X_baseH, [targetH{l} ones(TsizeH,1)],'binomial', 'link', 'logit','weights',wetH);
265       temp =  targetH{l};
266       PfitH = glmval(thiH{l},FeaVectorH' , 'logit')>=0.5;
267       FNH{l} = sum(PfitH(logical(temp))==0)./sum(logical(temp));
268       FPH{l} = sum(PfitH(~logical(temp))==1)./sum(~logical(temp));
269       Pfit_baseH = glmval(thi_baseH{l},X_baseH , 'logit') >= 0.5;
270       FN_baseH{l} = sum(Pfit_baseH(logical(temp))==0)./sum(logical(temp));
271       FP_baseH{l} = sum(Pfit_baseH(~logical(temp))==1)./sum(~logical(temp));
272       if TsizeV ~=0
273          [thiV{l}] = glmfit(FeaVectorV', [targetV{l} ones(TsizeV,1)],'binomial', 'link', 'logit','weights',wetH);
274          X_baseV = ones(TsizeV,1);
275          [thi_baseV{l}] = glmfit(X_baseV, [targetV{l} ones(TsizeV,1)],'binomial', 'link', 'logit','weights',wetH);
276          temp =  targetV{l};
277          PfitV = glmval(thiV{l},FeaVectorV' , 'logit')>=0.5;
278          FNV{l} = sum(PfitV(logical(temp))==0)./sum(logical(temp));
279          FPV{l} = sum(PfitV(~logical(temp))==1)./sum(~logical(temp));
280          Pfit_baseV = glmval(thi_baseV{l},X_baseV , 'logit') >= 0.5;
281          FN_baseV{l} = sum(Pfit_baseV(logical(temp))==0)./sum(logical(temp));
282          FP_baseV{l} = sum(Pfit_baseV(~logical(temp))==1)./sum(~logical(temp));
283       else
284          [thiV{l}] = [thiH{l}];
285          [thi_baseV{l}] = [thi_baseH{l}];
286          targetV{l} = targetH{l};
287          X_baseV = ones(TsizeH,1);
288          PfitV = PfitH;
289          FPV{l} = FPH{l};
290          FNV{l} = FNH{l};
291          Pfit_baseV = Pfit_baseH;
292          FP_baseV{l} = FP_baseH{l};
293          FN_baseV{l} = FN_baseH{l};
294       end
295
296       % calculate error
297
298    elseif strcmp(LearnAlg,'GDA')
299    elseif strcmp(LearnAlg,'SVM')
300    end
301    % end learning ==============================================================================
302l = l +1;
303DateStamp = date;
304save([ScratchDataFolder '/../learned_parameter/DisCts_' ImgFolder '_' LearnAlg ...
305      '_Nonsky' num2str(SkyExclude) '_Log' num2str(logScale) ...
306      '_Near' num2str(LearnNear) '_WeiBatNu' num2str(WeiBatchNumber) ...
307      '_' AbsFeaType '_AbsFeaDate' AbsFeaDate '_' ...
308      HistFeaType '_HistFeaDate' HistFeaDate '_LearnDate' DateStamp '.mat'],...
309      'thiH','thiV','thi_baseH','thi_baseH','FPH','FPV','FNH','FNV','FP_baseH','FP_baseV','FN_baseH','FN_baseV');
310end
311
312DateStamp = date;
313save([ScratchDataFolder '/../learned_parameter/DisCts_' ImgFolder '_' LearnAlg ...
314      '_Nonsky' num2str(SkyExclude) '_Log' num2str(logScale) ...
315      '_Near' num2str(LearnNear) '_WeiBatNu' num2str(WeiBatchNumber) ...
316      '_' AbsFeaType '_AbsFeaDate' AbsFeaDate '_' ...
317      HistFeaType '_HistFeaDate' HistFeaDate '_LearnDate' DateStamp '.mat'],...
318      'thiH','thiV','thi_baseH','thi_baseH','FPH','FPV','FNH','FNV','FP_baseH','FP_baseV','FN_baseH','FN_baseV');
319%if logScale == 1
320%   if SkyExclude == 1
321%      save([ScratchDataFolder '/../learned_parameter/DisCts_' ImgFolder '_' LearnAlg '_Nonsky_WeiBatNu' num2str(WeiBatchNumber) '_' AbsFeaType '_AbsFeaDate' AbsFeaDate '_' HistFeaType '_HistFeaDate' HistFeaDate '_LearnDate' DateStamp '.mat'],'thiH','thiV','thi_baseH','thi_baseH','errorH','errorV','error_baseH','error_baseV','learnRatioH','learnRatioV');
322%   else
323%      save([ScratchDataFolder '/../learned_parameter/DisCts_' ImgFolder '_' LearnAlg '_WeiBatNu' num2str(WeiBatchNumber) '_' AbsFeaType '_AbsFeaDate' AbsFeaDate '_' HistFeaType '_HistFeaDate' HistFeaDate '_LearnDate' DateStamp '.mat'],'thiH','thiV','thi_baseH','thi_baseH','errorH','errorV','error_baseH','error_baseV','learnRatioH','learnRatioV');
324%   end
325%else
326%   if SkyExclude == 1
327%      save([ScratchDataFolder '/../learned_parameter/DisCts_' ImgFolder '_' LearnAlg '_Nonsky_WeiBatNu' num2str(WeiBatchNumber) '_' AbsFeaType '_AbsFeaDate' AbsFeaDate '_' HistFeaType '_HistFeaDate' HistFeaDate '_LearnDate' DateStamp '_linear.mat'],'thiH','thiV','thi_baseH','thi_baseH','errorH','errorV','error_baseH','error_baseV','learnRatioH','learnRatioV');
328%   else
329%      save([ScratchDataFolder '/../learned_parameter/DisCts_' ImgFolder '_' LearnAlg '_WeiBatNu' num2str(WeiBatchNumber) '_' AbsFeaType '_AbsFeaDate' AbsFeaDate '_' HistFeaType '_HistFeaDate' HistFeaDate '_LearnDate' DateStamp '_linear.mat'],'thiH','thiV','thi_baseH','thi_baseH','errorH','errorV','error_baseH','error_baseV','learnRatioH','learnRatioV');
330%   end
331%end
Note: See TracBrowser for help on using the repository browser.