ciklum & nvidia Deep Learning Practice Beyond the Hype Nov 13, Frankfurt am Main Register now



Improve the accuracy of your predictions Image Recognition & Object Detection for The Nature Conservancy

Image Classification Icebergs Ciklum R&D Team Built the Image Classification Algorithm to Identify Threatening Icebergs

DETECT OBJECTS IN A NEARLY HUMAN-LEVEL PERFORMANCE AI model to automatically detect fruits from an image for SeeTree

Save time and money by streamlining the online buying process Car Image Masking With Convolutional Neural Networks

LET’S TALKGet in touch with our engineers to find out how you can automate routine tasks and improve accuracy of predictions with Deep Learning

Object Detection

Image Recognition & Object Detection for The Nature Conservancy


The Nature Conservancy is one of the world’s leading conservation organisations, working in more than 70 countries to protect ecologically important lands and waters for nature and people. In the Western and Central Pacific, where 60% of the world’s tuna is caught, illegal, unreported, and unregulated fishing is threatening marine ecosystems, and global seafood supplies. The Nature Conservancy wanted to develop a camera-based solution for automatic detection and classification of fish species caught by the commercial fishing fleet. To find the solution they summoned the global tech community to participate in a Fisheries Monitoring contest at Kaggle.


Ciklum R&D Engineering team received the training dataset of about 3,700 photos. The team needed to build an algorithm to identify a certain fish type. Traditional algorithms would not be effective enough for the required solution, hence Ciklum engineers used innovative object detection methods based on convolutional neural networks.


Ciklum R&D team developed a series of algorithms to automatically detect and classify different species of tuna, sharks and more that fishing boats catch, to accelerate the video review process. Fish detection algorithm – localization precision is 95%, recall – 80% Classification algorithm for different species of fish on a boat with 93% accuracy

SeeTree is a leading company in the agritech field with a focus on tree farming. One of the applications SeeTree is targeting is fruit counting, which includes automatic detection of fruits on the trees from an image. SeeTree partnered with Ciklum to approach this task through deep learning which is considered the state-of-the-art in tackling such object detection tasks.

In 2 months Ciklum R&D team built the AI prototype that can outperform human in accuracy, especially on detection of occluded fruits.



Statoil is an international energy company partnered with C-CORE that has been using satellite data for 30+ years to build a computer vision based iceberg surveillance system. C-CORE and Statoil wanted to find better ways to locate icebergs before they drift near oil and gas infrastructure and needed to approach the problem from a different perspective. The major challenge was to build an algorithm to automatically identify whether a remotely sensed target is an iceberg or not. The algorithm had to be extremely accurate because lives and billions of dollars in energy infrastructure at stake.


To build a model, the Ciklum R&D team used multiple deep learning models ensembled together. Ensemble techniques use two or more learning algorithms to get better predictive performance than could be obtained from any of the constituent learning algorithms alone. The output of such a system is better and more precise than the output of each method separately. The final model included 40 models. Each was determined beforehand according to the voting principle on the basis of 5 submodels:


The final result was evaluated using log loss metrics. This metric measures the accuracy of the model where the prediction input is a probability value between 0 and 1. The goal of our machine learning models is to minimize this value. A perfect model would have a log loss of 0. The system developed by the Ciklum R&D team defined the iceberg with a log loss of 0.1310 Read more details in our blog

Car Image Masking

Car Image Masking With Convolutional Neural Networks


Carvana is an online marketplace for buying used vehicles. It provides a fully automated service scheme that resembles a coin-operated vending machine for cars. Carvana has a custom rotating photo studio that automatically captures and processes 16 standard images of each vehicle in their inventory. Carvana takes high-quality photos, but bright reflections and cars with colors similar to the background cause automation errors, which requires a skilled photo editor to change. To manually edit 16 projections of thousands of individual cars, the company would have to hire an army of designers, which would be a time-consuming and expensive process. Customers had to rely on blurry pictures with little information. Carvana urgently needed an algorithm to separate the car from the background as a professional designer would. The company hosted a competition on Kaggle with $25,000 prize, 735 teams, and two months to find a solution.

Auto editing


Ciklum R&D team had to develop an algorithm to automatically generate masks for cars in images. The algorithm has to predict whether each pixel of the picture belongs to the class of car or the background. For the training set, a GIF file with 2.5K images was provided, containing the manual cutout mask for each image with pixel intensity values of 0 and 1. The testing set was 100,064 images in .jpg format with 1280 x 1918 resolution. The car and the background in the image had to be colored at the pixel level. The input was a JPG file and the output, a binary mask. There are several popular models for semantic segmentation in recent deep learning literature such as SegNet, FCN, E-NEt, U-Net, and this year’s state-of-the-art models, such as PSPNet and LinkNet. After some initial experimentation, SegNet and U-Net were used as a base to build a custom architecture. A model was built from the best results on the basis of weighted values. Depending on internal validation, the values were summed up for the next validation.


The evaluation metric was a mean Dice coefficient, which demonstrated how two multitudes are overlapped. The higher the Dice coefficient, the better (max 1). The algorithm demonstrated a mean Dice coefficient of 0.997 for differentiating the cars from the background, which was much better and faster than manual design work. How the algorithm works:

algorithm works

With an efficient algorithm, Carvana could now build trust with consumers by providing better images, streamline the online buying process, and most importantly, save time and money.