luciabaldassini / grasping_detection_system

A deep learning approach to predict the grasping areas of objects

Grasping Detection System

Introduction

This papers investigates several approaches for the detection of graspable areas of an object. In particular, two Resnet-50 and two ResNet-18 models are trained with different configurations:

1. Pretrained ResNet-18 with a linear activation function
2. Pretrained ResNet-18 with a sigmoid activation function and re-scaling of the labels and output vales
3. Pretrained ResNet-50 with a sigmoid activation function and re-scaling of the labels and output vales
4. Non-pretrained ResNet-50 with a sigmoid activation function and re-scaling of the labels and output vales.

Dataset

The models are trained on the Cornell Grasping Dataset, which can be downloaded from http://pr.cs.cornell.edu/grasping/rect_data/data.php

Results

The result show that all model achieve an accuracy of around 90%. Out of the four configurations, the ones re-scaling the labels and the output and using a sigmoid activation function (models 2, 3 and 4) resulted rectangles that are better suited for grasping, in terms of orientation, compared to model 1, which did not use such approach.

grasping_detection_system's Issues

Documentation