Congratulations to Guoxiang Zhang and Tiebiao Zhao for winning NVidiab GPU grant (Titian X Pascal)

March 31, 2017

Congratulations to Guoxiang Zhang and Tiebiao Zhao for winning NVidia GPU grant (Titian X Pascal)

(Performace comparisons)


From: NVIDIA GPU Grant Program []
Sent: Wednesday, March 22, 2017 1:34 PM
To: YangQuan Chen <>
Subject: NVIDIA GPU Grant Approved - YangQuan Chen

Hi YangQuan,


We have reviewed your NVIDIA GPU Grant Request and we are happy support your work with the donation of (1) Titan X Pascal to support your research. We will ship via FedEx/DHL to the address listed in the form below and you can expect to receive within two weeks.  Please respond within 48 hours if you are unable to house the GPU, unable to clear customs or the address, email or phone number you have provided is incorrect.  If you are unable to house the GPU or the shipment is unable to be delivered we will not be able to re-ship.  Once the GPU has shipped you will receive a courtesy email from our shipping team with the tracking number.


We ship donated GPUs with a commercial invoice per shipping regulations. The price listed on this commercial invoice is the donation value not the price you would purchase online or from a vendor at MSRP price. Our vendors decide their MSRP pricing. We also ship so we pay all fees if the country will allow it.


Please ensure that the GPU is powered properly as per the information found here:


If this hardware donation results in any publications or reports, we kindly ask that you acknowledge “NVIDIA Corporation”.  This is one way we track and justify our programs and adding “NVIDIA Corporation” to your acknowledgement will help us more easily track. We trust your judgment in how you acknowledge NVIDIA Corporation, however here are a couple samples:  "The Titan X Pascal used for this research was donated by the NVIDIA Corporation."  or “We gratefully acknowledge the support of NVIDIA Corporation with the donation of the Titan X Pascal GPU used for this research.” You are also encouraged to acknowledge NVIDIA Corporation in any contacts with the news media or in general articles.


This equipment donation will come as an unrestricted gift to support your research, however it is not for resale.  As part of the GPU Grant Program, you will receive occasional newsletters and requests to complete surveys.  While you are free to opt out of the newsletter at any time, we ask that you complete the infrequent surveys. These are important for tracking and justifying our academic programs, and cover topics such as updates on your GPU related courses, publications and research progress.


All the best, 


Academic Programs Team



First Name


Last Name



Associate Professor


University of California, Merced

Shipping Address (preferably a University Address) 1

5200 North Lake Road

Shipping Address (preferably a University Address) 2


Shipping Address (preferably a University Address) 3





NALA (North America, Latin America)







Tax ID


Zip / Postal Code


Phone Number


Email Address

Research Website

Which libraries do you use or interested in using?

cuBLAS Complete BLAS Library, cuSPARSE Sparse Matrix Library, CHOLMOD

Research Domain/Field of Interest

Computational Photography, Computer Vision & Machine Vision, Virtual Reality

Programming interfaces/language solutions used for GPU acceleration

CUDA C++, Matlab, OpenGL / GLSL



Statement of Proposed Research

GPU Grant Proposal PI: Prof. YangQuan Chen; Univ. of California, Merced. Email:; iPhone: 209-3862958 Projects description: Our research group has two research projects that are related to general purpose graphics processing units (GPGPU) or convolutional neural network (CNN). 1) Exploring beneficial use of fractional calculus in 3D cave reconstruction and mapping using small low-cost UAVs as robotic co-archaeologists This project seeks to advance archaeological methods using drone technology, and to explore beneficial use of fractional calculus [1-7] in 3D reconstruction and mapping methods. It aims to develop drones capable of functioning as “Co-Archaeologists” that can map large caves and enter dangerous or hard to reach spaces. Its goal is to produce 3D models and semantic maps, which can both reduce the amount of labor and risk that cave archaeologists need to spend and take, and improve the quality of data and information acquired. Using RGB-D data collected by drones, we will be able to produce accurate, well textured 3D models with proper lighting and co-supervised by human archaeologists. It is still a challenge problem to reconstruct 3D models of indoor scenes, especially when the indoor environment is a dark cave. First, RGB-D cameras have a limited field of view and working range, which will bring in two problems: 1) More 3D model pieces need to be put together; 2) Each view only covers a small portion of the scene with limited information, which makes tracking prone to failure, especially when some areas of environment do not have much shape and color variation. Second, cave environment usually contains sophisticated geometric structures and objects, and must be scanned from complex camera trajectories for better coverage, which means one place can be observed multiple times from different view angle. Existing 3D reconstruction systems, including both online ones: simultaneous localization and mapping (SLAM), and offline approaches, cannot provide a solution that is reliable enough for our purpose. We plan to add human-in-the loop to address this reliability issue first, since we have observed huge reconstruction quality improvement by only add very few human interference, specifically, by closing a few loops when rescan the same region several times. We choose to add human in the loop because current loop closure detection methods in SLAM methods cannot provide satisfactory results, which lead to the second stage of this project: improve current loop closure methods. In this part, we will develop algorithms that can take multiple observations as input as well as actively control where to observe, because current methods process keyframes separately as a single observation, and only use low level computer vision technique, such as bag of word and randomized trees. More importantly, after generating high fidelity 3D models. CNN based methods will be explored to parse them, including 3D segmentation and recognition. This step is crucial, because it can help convert 3D models to high level abstracted 3D semantic maps, which is important for archaeologists to analyze. Instead of using existing point cloud based 3D segmentation and recognition or develop a CNN model that work on 3D space directly, we plan to use a way that mimic how human think and understand surrounding 3D environment: segmentation and recognition will still happen in 2D image space, but 3D model will be used as a proxy to connect multiple observation of the environments. This makes our method different from CNN or recurrent neural network (RNN) methods, in which we believe is one of the key components to next boost of visual data understanding. This technology has much broader impacts than a specific project. It can create new opportunities in the field of computer vision and robotics, especially in consumer robots, because the outcome of this project can be used to help robots to understand and remember world in 3D by only using consumer-level hardware and sensors. It sure will change the way archaeologists work as well. Archaeological research requires heavy data collection, yet recording and bookkeeping caves are a slow and tedious process. Our collaborator Prof. Holley Moyes is a leading specialist in cave archeology and ancient religions. Over her career, she has worked to develop new mapping methodologies, but finds that caves in Belize are being looted more rapidly than they can be investigated. Typically, archaeological teams will visit a site and begin to record it in one year, but when they come back to finish data collection it has been looted, artifacts stolen, architecture destroyed and the archaeological record disturbed. Therefore, archaeologists need a faster, more efficient method of surveying and recording the sites. This project can produce a working system that satisfies their need. 2) Remote-sensing image processing using deep learning Recently with the fast development of unmanned aerial vehicles (UAVs) and small payloads, there is a great revolution which makes remote sensing 2.0 happen. In contrast to traditional remote sensing using satellites, manned aircrafts, UAV-based remote sensing provides remote sensing with higher spatial resolution, more affordable cost, more flexible operation and more robustness to weather such as clouds. An ideal application of this technologies is crop monitoring in agriculture, including water stress detection, nutrient stress detection,integrated pest management (IPM), etc. Water stress detection helps farmers optimize the irrigation schedule and give crops water with right amount, right place and right time (3R). Nutrient stress detection allows growers to fine tune the application of nutrients according to 3R principles. Precise monitoring of crops growing status and pest infection is necessary in IPM in order to minimize use of chemicals and protect the environment. However, all these applications of UAV-based remote sensing lie on efficient image processing to convert big data to actionable information. Different from traditional low spatial resolution images that limit the monitoring scale to field level, UAV-based higher resolution images make it possible to tell the information in the scale as fine as tree levels or crop levels. Coming along with the benefits, there is a big challenge of processing images efficiently and extract accurate information regarding crops or plants. A typical example is how to differentiate crops from soil and shade, and to classify the illuminating part of canopies, which is the first step for crop-level or plan-level stress detection. Another example is in IPM practice, weed classification and quantification is necessary to determine the management tactics. To deal with these problems, it requires to extract features robust to lighting condition, image resolutions, crop growing stages and species. Instead of traditional image processing methods, deep learning tools promise to handle these challenges by learning all kinds of features automatically. Located in the central valley, California, MESA lab started research of drone’s applications in agriculture since 2014. So far we have been working on many crops, such as almond, pomegranate, cherry, onion, cantaloupe, etc. The research questions are stress detection, yield prediction, and weed management. . More specifically, we have the following research questions: 1) How to extract tree canopy [8] in the images more accurately and easily. By focusing on the region of canopies of almonds and pomegranate, we will have better measurement accuracy on either stress detection or yield prediction 2) How to classify flowers of cherry trees in the images from soil, branches and shade. By quantifying the size and number of flowers, we are interested in predicting yield and harvest time. 3) How to classify weeds in the crop field such as cantaloupe and onions. By quantifying the size and species of weeds, we want to know what is the best management tactics do conduct weed control. 4) How to classify cantaloup fruit in the field. This will help yield estimation. How we will use the GPU We plan to use this GPU in both two projects. In the first project, current offline 3D reconstruction methods take several hours to get results on a modern CPU. We plan to utilize massive parallel processing power of GPU to accelerate this process. First, in iterative closest point (ICP), we will use OpenGL to render 3D models, and then use GPU shaders to get predicted view of models. Then, projective frame-to model data association, instead of K nearest neighbor search, will be used to find correspondences for least square optimization, which can greatly boost the performance. Second, the most time consuming part of this process is highly parallelizable. There are hundreds or thousands of pairs of 3D models to be matched together, while each pair is independent from all the other pairs. Even inside matching one pair of 3D models, thousands of point correspondences independently contribute to final residual and Jacobian. This makes it possible to use CUDA programming to distribute workload to different CUDA cores. We expect to make a system that work in real time, while still has result comparable to the best offline method, so that we will have both 3D reconstruction quality and online feedback, which is beneficial for both researchers and final end users. Second, in order to generate 3D semantic maps, we will use deep neural networks libraries, such as cuDNN and Caffe, to parse each RGB frame as an image to have a frame level coarse segmentation result. Then, since we have multiple observations of a same place on the 3D model, and frame level segmentation result can be back projected to 3D model, we will develop mathematical algorithms to aggregate these observation to get a semantic label for each object and place in the 3D model with high confidence. 2) In the second project, we will train and fine-tune different CNN based object detection models, such as faster R-CNN, You Only Look Once (YOLO), etc., using Caffe on our labeled data, and then benchmark these models in crops detection, classification tasks on our test images. Then we will make changes to the best or most suitable model to make further adaptations that take specific prior of our problem to further improve performance. References Cited: [1] Chen, D., Chen, Y. and Xue, D., 2012. 1-D and 2-D digital fractional-order Savitzky–Golay differentiator. Signal, Image and Video Processing, 6(3), pp.503-511. [2] Knight, J., Smith, B. and Chen, Y., 2014, September. An essay on unmanned aerial systems insurance and risk assessment. In Mechatronic and Embedded Systems and Applications (MESA), 2014 IEEE/ASME 10th International Conference on (pp. 1-6). IEEE. [3] Stark, Brandon, Brendan Smith, YangQuan Chen. 2013. A Guide for Selecting Small Unmanned Aerial Systems for Research-Centric Applications. The 2nd IFAC Workshop on Research, Education and Development of Unmanned Aerial Systems (RED-UAS, 2013. November 20-22, 2013, Compiegne, France. [4] Tricaud, Christophe and YangQuan Chen. 2012. Optimal Mobile Sensing and Actuation Policies in Cyber-physical Systems. Springer, NY. ISBN 978-1-4471-2261-6. [5] Chen, Y.Q. and Moore, K.L., 2002. Discretization schemes for fractional-order differentiators and integrators. IEEE Transactions on Circuits and Systems I: Fundamental Theory and Applications, 49(3), pp.363-367. [6] Dali Chen, Dingyu Xue, YangQuan Chen, 2012, May. “Fractional differentiation-based approach for robust image edge detection” (#046) Proceedings of 2012 Fractional Derivative and Applications (FDA2012), Nanjing, China. [7] Tian, D., Xue, D., Chen, D. and Sun, S., 2013, May. A fractional-order regulatory CV model for brain MR image segmentation. In 2013 25th Chinese Control and Decision Conference (CCDC) (pp. 37-40). IEEE. [8]T Zhao, M Cisneros, Q Yang, Y Zhang, Y Chen, 2016. Almond Canopy Detection and Segmentation Using Remote Sensing Data Drones. In 2016 13th International Conference on Precision Agriculture (ICPA), St. Louis, USA. [9]T Zhao, J Gonzalez, J Franzen, Q Yang, Y Chen, 2016. Melon Classification and Segmentation Using Low Cost Remote Sensing Data Drones. In 2016 13th International Conference on Precision Agriculture (ICPA), St. Louis, USA. [10]T Zhao, Z Wang, Y Chen, 2017. Melon yield prediction using small unmanned aerial vehicles. SPIE Commercial + Scientific Sensing and Imaging, Anaheim, USA.

Equipment Requested


Do you have equipment to house the donated GPUs?


CV or Bio


Additional Materials


Are you interested in sharing your work with the community?