Fixed mAP calculation for CRNN layer during training

README.md

@@ -8,7 +8,7 @@
 * [Explanations in issues](https://github.com/AlexeyAB/darknet/issues?q=is%3Aopen+is%3Aissue+label%3AExplanations)
 
 0.  [Improvements in this repository](#improvements-in-this-repository)
-1.  [How to use](#how-to-use)
+1.  [How to use](#how-to-use-on-the-command-line)
 2.  [How to compile on Linux](#how-to-compile-on-linux)
 3.  How to compile on Windows
     * [Using vcpkg](#how-to-compile-on-windows-using-vcpkg)

scripts/windows/otb_get_labels.sh

@@ -0,0 +1,50 @@
+#!/bin/bash
+ 
+dataset=$1
+w=$2
+h=$3
+
+# Parameters: Human3, CarScale, Human6, Biker
+#w=480
+#h=640
+IFS=','
+  
+export LC_NUMERIC="en_US.UTF-8"
+
+wd=`pwd`
+dataset_path="data/$dataset"
+
+class_id=0
+num=1
+
+mkdir data
+wget http://cvlab.hanyang.ac.kr/tracker_benchmark/seq/$dataset.zip
+
+unzip -o $dataset.zip -d data
+
+sed -i.bak 's/\o11/,/g' $dataset_path/groundtruth_rect.txt
+sed -i.bak 's/\o11/,/g' $dataset_path/groundtruth_rect.txt
+dos2unix $dataset_path/groundtruth_rect.txt
+
+while read -r left right width height; do
+filename=$(printf "$dataset_path/img/%04d.txt" $num)
+#rm $filename.txt
+echo "$class_id " > $filename
+printf "%.5f " "$(($((left + width/2))  * 100000 / $w))e-5"   >> $filename
+printf "%.5f " "$(($((right + height/2))  * 100000 / $h))e-5"   >> $filename
+printf "%.5f " "$(($((width))  * 100000 / $w))e-5"   >> $filename
+printf "%.5f " "$(($((height))  * 100000 / $h))e-5"   >> $filename
+num=$((num + 1))
+done < $dataset_path/groundtruth_rect.txt
+
+echo "$dataset" > $dataset_path/otb.names
+ 
+
+find $dataset_path/img -name \*.jpg > data/$dataset/train.txt
+
+echo "classes = 1" > data/otb_$dataset.data
+echo "train = data/$dataset/train.txt" >> data/otb_$dataset.data
+echo "valid = data/$dataset/train.txt" >> data/otb_$dataset.data
+echo "names = $dataset_path/otb.names" >> data/otb_$dataset.data
+echo "backup = backup/" >> data/otb_$dataset.data
+echo "results= results/" >> data/otb_$dataset.data

scripts/windows/win_get_otb_datasets.cmd

@@ -0,0 +1,18 @@
+echo Run install_cygwin.cmd before:
+
+rem http://cvlab.hanyang.ac.kr/tracker_benchmark/datasets.html
+
+c:\cygwin64\bin\bash -l -c "cd %CD:\=/%/; echo $PWD"
+
+c:\cygwin64\bin\dos2unix "%CD:\=/%/otb_get_labels.sh"
+
+
+c:\cygwin64\bin\bash -l -c "cd %CD:\=/%/; %CD:\=/%/otb_get_labels.sh Suv 320 240"
+
+c:\cygwin64\bin\bash -l -c "cd %CD:\=/%/; %CD:\=/%/otb_get_labels.sh Liquor 640 480"
+
+c:\cygwin64\bin\bash -l -c "cd %CD:\=/%/; %CD:\=/%/otb_get_labels.sh Freeman4 360 240"
+
+c:\cygwin64\bin\bash -l -c "cd %CD:\=/%/; %CD:\=/%/otb_get_labels.sh Human3 480 640"
+
+pause

scripts/windows/win_get_otb_human3.cmd

@@ -1,22 +0,0 @@
-echo Run install_cygwin.cmd before:
-
-rem http://cvlab.hanyang.ac.kr/tracker_benchmark/datasets.html
-
-c:\cygwin64\bin\bash -l -c "cd %CD:\=/%/; echo $PWD"
-
-
-c:\cygwin64\bin\wget http://cvlab.hanyang.ac.kr/tracker_benchmark/seq/Human3.zip
-
-c:\cygwin64\bin\unzip -o "%CD:\=/%/Human3.zip"
-
-
-
-c:\cygwin64\bin\dos2unix "%CD:\=/%/windows_otb_get_labels.sh"
-
-c:\cygwin64\bin\bash -l -c "cd %CD:\=/%/; %CD:\=/%/windows_otb_get_labels.sh"
-
-rem c:\cygwin64\bin\find "%CD:\=/%/img" -name \*.jpg > otb_train.txt
-
-
-
-pause

scripts/windows/windows_otb_get_labels.sh

@@ -1,28 +0,0 @@
-#!/bin/bash
-
-export LC_NUMERIC="en_US.UTF-8"
-
-wd=`pwd`
-IFS=','
-
-dataset="Human3"
-class_id=0
-w=480
-h=640
-num=1
-
-while read -a line; do
-filename=$(printf "$dataset/img/%04d.txt" $num)
-#rm $filename.txt
-echo "$class_id " > $filename
-printf "%.5f " "$(($((line[0] + line[2]/2))  * 100000 / $w))e-5"   >> $filename
-printf "%.5f " "$(($((line[1] + line[3]/2))  * 100000 / $h))e-5"   >> $filename
-printf "%.5f " "$(($((line[2]))  * 100000 / $w))e-5"   >> $filename
-printf "%.5f " "$(($((line[3]))  * 100000 / $h))e-5"   >> $filename
-num=$((num + 1))
-done < $dataset/groundtruth_rect.txt
-
-echo "$dataset" > $dataset/otb.names
- 
-
-find $dataset/img -name \*.jpg > otb_train.txt

src/crnn_layer.c

@@ -263,7 +263,7 @@ void forward_crnn_layer_gpu(layer l, network_state state)
     s.train = state.train;
     s.workspace = state.workspace;
     s.net = state.net;
-    if(!state.train) s.index = state.index;  // use TC only for detection
+    //if(!state.train) s.index = state.index;  // don't use TC at all (especially without cuda_convert_f32_to_f16() )
     int i;
     layer input_layer = *(l.input_layer);
     layer self_layer = *(l.self_layer);