Learnings from some of the projects we’ve worked on
Mass Transit: Detecting masks and belongings of public commuters
Our client wanted to use AI to detect transit commuters from CC Camera footage (including fisheye lens) – their gender, posture, what they are carrying, whether they are wearing face masks, etc.
Tools used: Supervisely, CVAT, Darwin by v7labs
Challenges we faced:
The project had a very tight deadline with a large dataset with loads of different labels on each image. We had to train and onboard new workers at an unprecedented rate, keeping our promised quality in mind. The team sprinted day and night and delivered 99%+ accuracy in a quick three week turnaround.
Biomedicine: Project Neuron - Where’s the Axon?
Our client wanted to annotate 3D microscopic images of nuclei of neurons (with segmentation) and trace the dendrites with 3D lines. They evaluated 7 different vendors for the project… and we’re happy to say that we Abelling was chosen as the preferred workforce partner!
Tools used: WebKNOSSOS
Challenges we faced: The specialized nature & complexity of the project meant that typical annotators would not suffice. Thankfully, we were able to source & train subject-matter experts from university students of Life Science subjects and Doctors at a competitive rate.
The project was successful with 5 fold redundancy and we are proud to have contributed in development of the solution for the client.
Public Infrastructure: Repairing telephone poles in Japan
As part of a larger initiative for automating the maintenance of poles in Japan’s telephone grid, our client wanted to use computer vision to identify individual poles from their respective number plates.
Tools used: Supervisely
Challenges we faced: This was a particularly challenging task, mainly due to 3 issues:
- The number plates on the poles contained a mixture of the Kanji & Hiragana character sets used in Japanese writing – leading to added complexity for labellers whilst annotating
- Well into the project, the client’s requirements changed and the brief got revised – necessitating us to perform multiple iterations on the same data sets.
- Number of classes per image was very high, making identification per-label very time-consuming for our labellers
Thankfully, our project managers worked closely with the client over the project’s month-long duration; Thanks to their collaborative efforts we were able to overcome above mentioned challenges and complete the project OTIF (on-time in-full) successfully.
Sports: Athlete pose estimation for Alpine skiing
Posture is super important when speeding down the icy slopes of the Alps! We helped our client recognize 25-point pose estimation skeletons of in-action athletes across thousands of frames captured from a variety of video clips.
Tools used: Darwin by v7labs
Challenges we faced: This was when we were blown away by the skeleton-specific features from Darwin, the brilliant computer vision annotation tool from the folks in London. This shaved off hours from our project and allowed us to quickly turn around with the correct points in the right joints!
Analyzing public sentiment around COVID from Tweets & Facebook posts
As COVID-19 led to unprecedented uncertainty & chaos in the daily lives of billions around the globe, our client wanted to help Public Policy Makers, Governments and MNCS listen to what the masses were saying on social media regarding the pandemic – and the overall sentiment around the interventions which local authorities were taking in their respective geographies.
Their model was to synthesize thousands of tweets, posts and comments to assess overall public sentiment and uncover the broad themes & insights embedded within them. We were happy to help feed their model by labeling sample sets of such social media messages – all the while ensuring complete user anonymity and data privacy, ofcourse.
Tools used: labelstud.io
Challenges we faced: Often the same message, post or comment had multiple tonalities (e.g. negative/critical of government in the beginning, positive/hopeful by the conclusion). This made it quite difficult to assign static labels for overlapping classes in the same data.
Posts also could contain multiple languages (often using colloquial or slang) – which meant we needed to onboard labellers conversational in those specific languages in-order to properly perform our task.