Densely Supervised Grasp Detector (DSGD)
Authors: Umar Asif, Jianbin Tang, Stefan Harrer8085-8093
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Evaluation on a multi-object dataset and real-world robotic grasping experiments show that DSGD produces highly stable grasps on a set of unseen objects in new environments. It achieves 97% grasp detection accuracy and 90% robotic grasping success rate with real-time inference speed. |
| Researcher Affiliation | Industry | Umar Asif IBM Research Australia umarasif@au1.ibm.com Jianbin Tang IBM Research Australia jbtang@au1.ibm.com Stefan Harrer IBM Research Australia sharrer@au1.ibm.com |
| Pseudocode | No | The paper describes the system architecture and components but does not provide any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper provides links to experiment videos but does not explicitly state that the source code for DSGD is available or provide a link to a code repository. |
| Open Datasets | Yes | We evaluated DSGD for grasp detection on the popular Cornell grasp dataset (Lenz, Lee, and Saxena 2015), which contains 885 RGB-D images of 240 objects. The ground-truth is available in the form of grasp-rectangles. We also evaluated DSGD for multi-object grasp detection in new environments. For this, we used the multi-object dataset of (Asif, Tang, and Harrer 2018b) which consists of 6896 RGB-D images of indoor scenes containing multiple objects placed in different locations and orientations. |
| Dataset Splits | No | The paper mentions using 'object-wise splitting criteria' to split object instances randomly into 'train and test subsets', but it does not explicitly provide details for a separate validation split or specific percentages/counts for these splits, beyond general methodology. |
| Hardware Specification | Yes | Training was performed using ADAM optimizer and data parallelism on four Nvidia Tesla K80 GPU devices. |
| Software Dependencies | No | Our implementation is based on the framework of Torch library (Paszke et al. 2017). |
| Experiment Setup | Yes | Specifically, we initialized the weights of the base network with the weights pre-trained on Image Net. For the Dense blocks (4-7), the fully connected layers of GGPN, GEN, SRN, and RGPN, and the fully convolutional layers of PGN, we initialized the weights from zero-mean Gaussian distributions (standard deviation set to 0.01, biases set to 0), and trained the networks using the loss functions in Eq. 5, Eq. 6, Eq. 9, Eq. 12, and Eq. 13, respectively for 150 epochs. The starting learning rate was set to 0.01 and divided by 10 at 50% and 75% of the total number of epochs. The parameter decay was set to 0.0005 on the weights and biases. Our implementation is based on the framework of Torch library (Paszke et al. 2017). Training was performed using ADAM optimizer and data parallelism on four Nvidia Tesla K80 GPU devices. For grasp selection during inference, DSGD selects the most confident region-level grasp if its confidence score is greater than a confidence threshold (δrgn), otherwise DSGD switches to the PGN branch and selects the most confident pixel-level grasp. If the most confident pixel-level grasp has a confidence score less than δpgn, DSGD switches to the GGN branch and selects the global grasp as the output. Experimentally, we found that δrgn = 0.95 and δpgn = 0.90 produced the best grasp detection results. |