We are looking to hire a freelancer to design the complete pipeline for the following problem statement:
Hardware to be used:
Problem Statement and Solution required:
We feed one clothing at a time on the conveyor belt. We have a camera installed at top of the conveyor belt. The ultimate aim is to extract a dominant color of each piece of cloth on the live stream. We want to implement this solution on Jetson Nano using their DeepStream pipeline.
We suggest following framework:
- The live stream should do ‘object detection’ and tracking of each individual cloth.
- After detection and unique tracking, we should perform ‘segmentation’.
- We should then find out the dominant color from the ‘segmented’ region of interest.
Data that will be provided by us:
- Over 6,000 images annotated for ‘object detection’.
- Over 6,000 images annotated for ‘segmentation’.
- For better training the annotated data can be increased by image augmentation, both for the ‘detection’ and ‘segmentation’. Our image data is currently on CVAT.
- Since we want to use Jetson Nano, we suggest we implement the complete pipeline using DeepStream.
- We suggest that we use Yolo v5 for the ‘detection’ (or any other model with high accuracy and less computational cost which is compatible with DeepStream).
- We should find out the most suitable ‘segmentation’ model compatible with Jetson Nano (and DeepStream).
- Since we want to do live ‘detection’ and ‘segmentation’ of clothes on conveyor belt, we suggest that we use tracking module available in DeepStream. The tracking module of DeepStream gives a unique ID to object during the live detection. Since, the ‘segmentation’ models are generally computationally expensive, therefore, it is recommended that a cloth with a ‘unique id’ is ‘segmented’ only once.
- Once the clothes are segmented then we can apply a filter on the segmented region to make the image blurry and then use any library (amongst many libraries available on GitHub) to extract the most dominant color (RGB Value) from the region-of-interest (that is the segmented region of the cloth).
We are looking for someone who can implement this project start-to-end including:
- Implementing DeepStream Pipeline
- Object Detection – Augment the existing annotated images (over 6000 images), train the model Yolo v5 (or any other model that works with DeepStream with high accuracy and FPS).
- Segmentation– Augment the existing annotated images (over 6000 images), train the Segmentation model that works with DeepStream with low computational cost.
- Convert object detection and segmentation models to Tensor RT or RTX (or any other solution) to make them work with DeepStream with live camera feed.
- Implement tracking on DeepStream.
- Dominant Color Extraction - The detected cloth with a unique ID should be segmented once and then dominant color from the segmented region should be extracted (a lot of codes are available on Github for dominant color detection).
Remote access to the Jetson device will be provided to you, as and when required.
Kindly provide your best quote and the timeframe required for the same and feel free to reach out to us for further details about the project/models/dataset.