Fast person detection

Intro
In this tutorial we will demonstrate how easy it is to build a production level model for person detection inside Supervisely without coding. At the end of this tutorial we will get both:
A huge training dataset for "person" detection task
Two custom state of the art models that are very fast and pretty accurate ("YOLO V3" and "SSD MobileNet V2")
Then we can download resulting models, get the source codes and embed them to our production environment (i.e. self-driving car, drone, robot etc.)
So, let's start.
Big Picture
First of all, we would like to display the entire pipeline from the high level perspective. It is good to see the big picture and understand the entire pipeline we are going to create.
Disclaimer: as you can see, we sequentially apply a few DTL queries. Of course, we can apply all the necessary transformations and augmentaions at the same time. We avoid doing that intentionally. It allows us to anylize the intermediate results visually and in terms of statistics.
Step 1. Upload Cityscapes
In our documentation you will find how to upload a public dataset (in Cityscapes format) to Supervisely. We will name this project as cityscapes_orig.
Uploaded project: 
​
All datasets inside the project: 
​
Supervisely automatically counts very useful statistics regarding images, objects, tags, and objects areas. Here are a few examples.
Statistics - project cityscapes_orig contains 5000 images in total: 
​
Statistics - objects number distribution in every project and in total: 
​
Step 2. DTL #1
This DTL takes original Cityscapes dataset and does the following:
1.
Layer #1 ("action": "data") merges two classes (rider and person) to the single one (person_poly) and all other classes to the single class unnec. We do not drop unnecessary classes because they are needed to make the correct rasterization.
2.
Layer #2 ("action": "rasterize") performs the correct rasterization and converts the class types to bitmap. Read more here.
3.
Layer #3 ("action": "drop_obj_by_class") drops class unnec_b with all its objects. As a result after this operation we will have objects of only the class person. Read more here (link1 and link2).
4.
Layer #4 ("action": "supervisely") saves results to the new project with the name cityscapes_rasterized_person.
1
[
2
  {
3
    "dst": "$sample",
4
    "src": [
5
      "cityscapes_orig/*"
6
    ],
7
    "action": "data",
8
    "settings": {
9
      "classes_mapping": {
10
        "rider": "person_poly",
11
        "person": "person_poly",
12
        "__other__": "unnec"
13
      }
14
    }
15
  },
16
  {
17
    "dst": "$f",
18
    "src": [
19
      "$sample"
20
    ],
21
    "action": "rasterize",
22
    "settings": {
23
      "classes_mapping": {
24
        "unnec": "unnec_b",
25
        "person_poly": "person"
26
      }
27
    }
28
  },
29
  {
30
    "dst": "$r",
31
    "src": [
32
      "$f"
33
    ],
34
    "action": "drop_obj_by_class",
35
    "settings": {
36
      "classes": [
37
        "unnec_b"
38
      ]
39
    }
40
  },
41
  {
42
    "dst": "cityscapes_rasterized_person",
43
    "src": [
44
      "$r"
45
    ],
46
    "action": "supervisely",
47
    "settings": {}
48
  }
49
]
Copied!
Here is the visualization of the current DTL query:
Step 3. DTL #2
In this DTL query we transform previous results.
1.
Layer #1 ("action": "data") merges the entire cityscapes_rasterized_person project and map person into person_b.
2.
Layer #2 ("action": "if") drops all images that have no objects.
3.
Layer #3 ("action": "bbox") converts objects of the class person_b from bitmaps to bounding boxes and renames this class as person.
4.
Layer #4 ("action": "resize") resizes image to the "height": 600.
5.
Layer #5 ("action": "objects_filter") drops objects of the class person that are less then 10px * 10px.
6.
Layer #6 ("action": "if") drops all images without any objects again.
7.
Layer #7 ("action": "supervisely") saves the results to the new project with the name city_person_filtered_boxes.
1
[
2
  {
3
    "dst": "$sample",
4
    "src": [
5
      "cityscapes_rasterized_person/*"
6
    ],
7
    "action": "data",
8
    "settings": {
9
      "classes_mapping": {
10
        "person": "person_b"
11
      }
12
    }
13
  },
14
  {
15
    "dst": [
16
      "$true",
17
      "null"
18
    ],
19
    "src": [
20
      "$sample"
21
    ],
22
    "action": "if",
23
    "settings": {
24
      "condition": {
25
        "min_objects_count": 1
26
      }
27
    }
28
  },
29
  {
30
    "dst": "$box",
31
    "src": [
32
      "$true"
33
    ],
34
    "action": "bbox",
35
    "settings": {
36
      "classes_mapping": {
37
        "person_b": "person"
38
      }
39
    }
40
  },
41
  {
42
    "dst": "$res",
43
    "src": [
44
      "$box"
45
    ],
46
    "action": "resize",
47
    "settings": {
48
      "width": -1,
49
      "height": 600,
50
      "aspect_ratio": {
51
        "keep": true
52
      }
53
    }
54
  },
55
  {
56
    "dst": "$filter",
57
    "src": [
58
      "$res"
59
    ],
60
    "action": "objects_filter",
61
    "settings": {
62
      "filter_by": {
63
        "polygon_sizes": {
64
          "action": "delete",
65
          "area_size": {
66
            "width": 10,
67
            "height": 10
68
          },
69
          "comparator": "less",
70
          "filtering_classes": [
71
            "person"
72
          ]
73
        }
74
      }
75
    }
76
  },
77
  {
78
    "dst": [
79
      "$true2",
80
      "null"
81
    ],
82
    "src": [
83
      "$filter"
84
    ],
85
    "action": "if",
86
    "settings": {
87
      "condition": {
88
        "min_objects_count": 1
89
      }
90
    }
91
  },
92
  {
93
    "dst": "city_person_filtered_boxes",
94
    "src": [
95
      "$true2"
96
    ],
97
    "action": "supervisely",
98
    "settings": {}
99
  }
100
]
Copied!
Here is the visualization of the current DTL query:
Step 4. Upload Mapillary
In our documentaion you will find how to upload a public dataset (in Mapillary format) to Supervisely. We will name this project as mapillary.
You can analize statisctics as we showed in Step 1.
Step 5. DTL #3
In this DTL query we apply transformations that are very similar to Step 3.
1.
Layer #1 ("action": "data") merges classes (Person, Bicyclist, Other Rider and Motorcyclist) into a single class (person_b) and drops all other classes.
2.
Layer #2 ("action": "if") drops all images that have no objects.
3.
Layer #3 ("action": "bbox") convert objects of class person_b from bitmaps to bounding boxes and renames this class as person.
4.
Layer #4 ("action": "resize") resizes image to the "height": 600.
5.
Layer #5 ("action": "objects_filter") drops objects of class person that are less then 10px * 10px.
6.
Layer #6 ("action": "if") drops all images without any objects again.
7.
Layer #7 ("action": "supervisely") saves the results to the new project with name mapillary_filtered_boxes.
1
[
2
  {
3
    "dst": "$sample",
4
    "src": [
5
      "mapillary/*"
6
    ],
7
    "action": "data",
8
    "settings": {
9
      "classes_mapping": {
10
        "Person": "person_b",
11
        "Bicyclist": "person_b",
12
        "__other__": "__ignore__",
13
        "Other Rider": "person_b",
14
        "Motorcyclist": "person_b"
15
      }
16
    }
17
  },
18
  {
19
    "dst": [
20
      "$true",
21
      "null"
22
    ],
23
    "src": [
24
      "$sample"
25
    ],
26
    "action": "if",
27
    "settings": {
28
      "condition": {
29
        "min_objects_count": 1
30
      }
31
    }
32
  },
33
  {
34
    "dst": "$box",
35
    "src": [
36
      "$true"
37
    ],
38
    "action": "bbox",
39
    "settings": {
40
      "classes_mapping": {
41
        "person_b": "person"
42
      }
43
    }
44
  },
45
  {
46
    "dst": "$res",
47
    "src": [
48
      "$box"
49
    ],
50
    "action": "resize",
51
    "settings": {
52
      "width": -1,
53
      "height": 600,
54
      "aspect_ratio": {
55
        "keep": true
56
      }
57
    }
58
  },
59
  {
60
    "dst": "$filter",
61
    "src": [
62
      "$res"
63
    ],
64
    "action": "objects_filter",
65
    "settings": {
66
      "filter_by": {
67
        "polygon_sizes": {
68
          "action": "delete",
69
          "area_size": {
70
            "width": 10,
71
            "height": 10
72
          },
73
          "comparator": "less",
74
          "filtering_classes": [
75
            "person"
76
          ]
77
        }
78
      }
79
    }
80
  },
81
  {
82
    "dst": [
83
      "$true2",
84
      "null"
85
    ],
86
    "src": [
87
      "$filter"
88
    ],
89
    "action": "if",
90
    "settings": {
91
      "condition": {
92
        "min_objects_count": 1
93
      }
94
    }
95
  },
96
  {
97
    "dst": "mapillary_person_filtered_boxes",
98
    "src": [
99
      "$true2"
100
    ],
101
    "action": "supervisely",
102
    "settings": {}
103
  }
104
]
Copied!
Here is the visualization of current DTL query:
Step 6. DTL #4
This DTL query merges transformed Cityscapes project (city_person_filtered_boxes) and transformed Mapillary project (mapillary_person_filtered_boxes),
1.
Layer #1 ("action": "data") gets everything from the project mapillary_person_filtered_boxes and keeps the classes as they are.
2.
Layer #2 ("action": "data") gets everything from the project city_person_filtered_boxes and keeps classes as they are.
3.
Layer #3 ("action": "dummy") combines data from Layer #1 and Layer #2 into a single variable.
4.
Layer #4 ("action": "flip") applies a vertical flip to the images and their annotations.
5.
Layer #5 ("action": "crop") performs relatively big random crops (from 60% to 90% in width and height with respect to the image size)
6.
Layer #6 ("action": "if") drops all images without any objects of class person.
7.
Layer #7 ("action": "if") randomly split the data into two branches: first branch - 98% (will be tagged as train) and second branch - 2% (will be tagged as val)
8.
Layer #8 ("action": "tag") adds the tag train to all input images.
9.
Layer #9 ("action": "tag") add the tag val to all input images.
10.
Layer #10 ("action": "supervisely") saves results to the new project with name train_detector.
1
[
2
  {
3
    "dst": "$sample",
4
    "src": [
5
      "mapillary_person_filtered_boxes/*"
6
    ],
7
    "action": "data",
8
    "settings": {
9
      "classes_mapping": "default"
10
    }
11
  },
12
  {
13
    "dst": "$sample2",
14
    "src": [
15
      "city_person_filtered_boxes/*"
16
    ],
17
    "action": "data",
18
    "settings": {
19
      "classes_mapping": "default"
20
    }
21
  },
22
  {
23
    "dst": "$data",
24
    "src": [
25
      "$sample",
26
      "$sample2"
27
    ],
28
    "action": "dummy",
29
    "settings": {}
30
  },
31
  {
32
    "dst": "$vflip",
33
    "src": [
34
      "$data"
35
    ],
36
    "action": "flip",
37
    "settings": {
38
      "axis": "vertical"
39
    }
40
  },
41
  {
42
    "dst": "$crop",
43
    "src": [
44
      "$data",
45
      "$vflip"
46
    ],
47
    "action": "crop",
48
    "settings": {
49
      "random_part": {
50
        "width": {
51
          "max_percent": 90,
52
          "min_percent": 60
53
        },
54
        "height": {
55
          "max_percent": 90,
56
          "min_percent": 60
57
        }
58
      }
59
    }
60
  },
61
  {
62
    "dst": [
63
      "$true_crop",
64
      "null"
65
    ],
66
    "src": [
67
      "$crop"
68
    ],
69
    "action": "if",
70
    "settings": {
71
      "condition": {
72
        "include_classes": [
73
          "person"
74
        ]
75
      }
76
    }
77
  },
78
  {
79
    "dst": [
80
      "$to_train",
81
      "$to_val"
82
    ],
83
    "src": [
84
      "$data",
85
      "$vflip",
86
      "$true_crop"
87
    ],
88
    "action": "if",
89
    "settings": {
90
      "condition": {
91
        "probability": 0.98
92
      }
93
    }
94
  },
95
  {
96
    "dst": "$train",
97
    "src": [
98
      "$to_train"
99
    ],
100
    "action": "tag",
101
    "settings": {
102
      "tag": "train",
103
      "action": "add"
104
    }
105
  },
106
  {
107
    "dst": "$val",
108
    "src": [
109
      "$to_val"
110
    ],
111
    "action": "tag",
112
    "settings": {
113
      "tag": "val",
114
      "action": "add"
115
    }
116
  },
117
  {
118
    "dst": "train_detector",
119
    "src": [
120
      "$train",
121
      "$val"
122
    ],
123
    "action": "supervisely",
124
    "settings": {}
125
  }
126
]
Copied!
Here is the visualization of the current DTL query:
Step 7. DTL #5
This is the final DTL query. As a result we will get the project train_detector_final that will be used for training (41615 images with 175088 objects).
1.
Layer #1 ("action": "data") gets everything from the project train_detector and keeps classes as they are.
2.
Layer #2 ("action": "objects_filter") drops objects of the class person that are less then 1px * 1px (drops "dummy" objects).
3.
Layer #3 ("action": "if") drops all images without any objects of class person.
4.
Layer #4 ("action": "supervisely") saves results to the new project with the name train_detector_final.
1
[
2
  {
3
    "dst": "$sample",
4
    "src": [
5
      "train_detector/*"
6
    ],
7
    "action": "data",
8
    "settings": {
9
      "classes_mapping": "default"
10
    }
11
  },
12
  {
13
    "dst": "$1",
14
    "src": [
15
      "$sample"
16
    ],
17
    "action": "objects_filter",
18
    "settings": {
19
      "filter_by": {
20
        "polygon_sizes": {
21
          "action": "delete",
22
          "area_size": {
23
            "width": 1,
24
            "height": 1
25
          },
26
          "comparator": "less",
27
          "filtering_classes": [
28
            "person"
29
          ]
30
        }
31
      }
32
    }
33
  },
34
  {
35
    "dst": [
36
      "$2",
37
      "null"
38
    ],
39
    "src": [
40
      "$1"
41
    ],
42
    "action": "if",
43
    "settings": {
44
      "condition": {
45
        "min_objects_count": 1
46
      }
47
    }
48
  },
49
  {
50
    "dst": "train_detector_final",
51
    "src": [
52
      "$2"
53
    ],
54
    "action": "supervisely",
55
    "settings": {}
56
  }
57
]
Copied!
Here is the visualization of the current DTL query:
Step 8. YOLO V3 training.
Basic step by step training guide is here. So it is the same for all models inside Supervisely.
Training for models YOLO V3 and SSD Mobilenet V2 will differ just in training configurations. Detailed information is here.
YOLO V3 weihts were initialized from the corresponding model from the Models list (it was pretrained on COCO).
For YOLO V3 we used this configuration:
1
{
2
  "lr": 0.0001,
3
  "epochs": 30,
4
  "batch_size": {
5
    "train": 64
6
  },
7
  "input_size": {
8
    "width": 448,
9
    "height": 448
10
  },
11
  "bn_momentum": 0.01,
12
  "gpu_devices": [
13
    0,
14
    1,
15
    2,
16
    3
17
  ],
18
  "data_workers": {
19
    "train": 3
20
  },
21
  "dataset_tags": {
22
    "train": "train"
23
  },
24
  "subdivisions": {
25
    "train": 16
26
  },
27
  "print_every_iter": 30,
28
  "weights_init_type": "transfer_learning",
29
  "enable_augmentations": true
30
}
Copied!
Training takes 6 hours on four GPUs. Here is the loss chart during training:
Step 9. SSD Mobilenet V2 training.
SSD Mobilenet V2 weihts were initialized from the orresponding model from the Models list (it was pretrained on COCO).
For SSD Mobilenet V2 we used this configuration:
1
{
2
  "lr": 0.0001,
3
  "epochs": 15,
4
  "val_every": 1,
5
  "batch_size": {
6
    "val": 12,
7
    "train": 12
8
  },
9
  "input_size": {
10
    "width": 600,
11
    "height": 600
12
  },
13
  "gpu_devices": [
14
    0,
15
    1,
16
    2,
17
    3
18
  ],
19
  "data_workers": {
20
    "val": 0,
21
    "train": 0
22
  },
23
  "dataset_tags": {
24
    "val": "val",
25
    "train": "train"
26
  },
27
  "weights_init_type": "transfer_learning",
28
  "allow_corrupted_samples": {
29
    "val": 0,
30
    "train": 0
31
  }
32
}
Copied!
Step 10. Download resulting models
"How to download model" guide is here.
Resources - Previous
Use Cases
Automatic roads segmentation
Last modified 1yr ago
Copy link
Contents
Intro
Big Picture
Step 1. Upload Cityscapes
Step 2. DTL #1
Step 3. DTL #2
Step 4. Upload Mapillary
Step 5. DTL #3
Step 6. DTL #4
Step 7. DTL #5
Step 8. YOLO V3 training.
Step 9. SSD Mobilenet V2 training.
Step 10. Download resulting models