Skip to content

Latest commit

 

History

History
144 lines (135 loc) · 6.07 KB

README.md

File metadata and controls

144 lines (135 loc) · 6.07 KB
Raw

Generate depth map from mono image in one request with CNNs (Convolutional neural networks)

REST API based on monodepth project (and other in future)

Setup

This section describes possible ways to deploy an application and app and deploy environment setup. There are two main ways to deploy an application:

  • Using docker-compose, which create isolated environment without dependency conflict (and also redis instance for caching)
  • Or just run run.py script, but then you have to deploy the redis instance yourself (more in cache section)

Docker

For build image use docker-compose build, it deploy project and automatically downloads all necessary dependencies and models for CNNs.
After build execute with docker-compose up or better use run-docker.sh script, which propose docker compose run mode (press Enter for execute in default mode, or type something and press Enter to run in demon mode)
Then you can access app on localhost:5000

Virtualenv

First, run get_models_monodepth.sh script, which downloads required models Use run-env.sh script which setup virtual environment, install depedncies and run app.

Configs

All environments variable (about flask variables) setted in config.env file and loading in config.py, where you can create own app config mode.

Cache

By default app support caching using Redis. For custom redis url set CACHE_REDIS_URL variable in config.env.
If you don't want to create redis instance or use docker, just change CACHE_TYPE to simple, and comment out CACHE_REDIS_URL = os.environ.get('CACHE_REDIS_URL') line in config.py (if you left CACHE_REDIS_URL variable in config.env, then it is not necessary)

CNNs setup

At the moment, there is only one CNN. When adding others, all logic will be described in cnn_name_bridge.py files

Monodepth

monodepth_bridge.py initializes all necessary models in advance, to avoid a long delay.
Model also has static input values for the height and width of the image (height - 256px and width - 512px), to maintain performance (this is especially noticeable when running on a CPU).
If GPU is used for calculations, change self_width and self_height in monodepth_bridge.py, or reinit models params for image. You can use method like this, for calculate optimal width and heigh and limit the size of the input image:

# image - BytesIO(image_bytes_arr)
def get_optimal_image_size(image_bytes):
    input_image = scipy.misc.imread(self.image_bytes, mode="RGB")
    # example for max 1248*960
    width, height, num_channels  = input_image.shape
    width = (1248 if(width > 1248) else (width - (width % 32)))
    height = (960 if(height > 960) else (height - (height % 32)))
    return width, height

API overview

Resource Method Description
/ GET Return list of available CNNs
/v1/cnns GET Return list of available CNNs
/v1/cnns/{cnn_name} GET Return list of available models for CNN
/v1/cnns/{cnn_name}/{model} POST Return predicted depth map (in png format) of image, using selected model

Endpoints

Available CNNs

Return array of available CNNs with external links to projects

  • URL: / or /v1/cnns
  • Method: GET
  • Success Response:
  • Error Response:
    • Code: 400
    • Content: {'message':'Bad Request error'}

Available models

Return array of available models for selected CNN

  • URL: /v1/cnns/{cnn_name}
  • Method: GET
  • URL params:
    • Required: cnn_name=string (received from GET /v1/cnns request in name filed)
  • Success Response:
    • Code: 200
    • Content: ["kitti", "cityscapes", "eigen"]
  • Error Response:
    • Code: 400
    • Content: {'message':'Bad Request error'}

Depth map

Return predicted depth map image in png format

  • URL: /v1/cnns/{cnn_name}/{model}
  • Method: POST
  • URL params:
  • Data params:
    • Image data with image form name, example below.
    • Supported image formats : png, jpg, jpeg
  • Request example: 
    POST /v1/cnns/monodepth/kitti
.....Here headings we are not interested in......
Content-Length: 2740
Content-Type: multipart/form-data;  boundary=----6PA4QswqtyuhfgxkTrZu0gW

----6PA4QswqtyuhfgxkTrZu0gW
Content-Disposition: form-data; name="image"; filename="test.jpg"
...........Here image data............
----6PA4QswqtyuhfgxkTrZu0gW
  • Success Response: 
    HTTP/1.0 200 OK
Content-Disposition: attachment; filename=cat.png
Content-Type: image/png
.....Image binary data........
  • Error Response:
    • Code: 400
    • Content: {'message':'Bad Request error'}

Examples

In this section described examples for POST request to generate depth map from image

cURL

curl -F 'image=@test.jpg' http://localhost:5000/v1/cnns/monodepth/kitti -o result.png

Python

import requests

url = 'http://localhost:5000/v1/cnns/monodepth/kitti'
files = {'image': open('test.jpg', 'rb')}

r = requests.post(url, files=files)
if r.status_code == 200:
    with open("result.png", 'wb') as f:
        f.write(r.content)
        f.close()