data_annotation

Data annotation with a QA review on every batch.

Images, video, text and documents, labelled to your guidelines. Every engagement starts with a paid pilot.

Images

Boxes, polygons and segmentation.

We label images with bounding boxes, polygons and segmentation masks. Each project needs clear definitions of the objects, classes and boundaries to label.

See labelled samples from our own photos or label them yourself

Video

Objects tracked frame by frame.

We label and track objects through video, frame by frame. The pilot is where we agree how to handle changing views, occlusion and objects whose identity is unclear.

Text

English and Bengali.

Text classification, named entity recognition and response rating in English and Bengali. We agree examples and rating rules with you before the main work starts.

Documents

Correct the OCR and label the fields you need.

OCR correction and document field extraction. The scope lists the fields required and how to treat unreadable or ambiguous material.

Pilot and QA

Agree the review standard on a small batch.

We sign an NDA before seeing your data. A separately scoped paid pilot comes before the main work, and every delivered batch gets a QA review. The pilot brief records the labelling rules, review criteria and output format for your sign-off.

Pricing depends on data volume, label complexity, review requirements and output format. We quote after scoping.

Labelled samples from our own photos of Dhaka.

Three images, three kinds of label. The photos are ours, taken by drone and from rooftops. Draft labels came from open models and were then corrected. These are demonstrations, not a client delivery, and a few small or hidden objects are left unlabelled.

Junction in old Dhaka seen from a rooftop, with boxes around people, cars, CNG auto-rickshaws, rickshaws, motorbikes, bicycles and a pickup — bounding_box
Boxes on a busy junction
Seven classes, including CNG auto-rickshaws and cycle rickshaws. The detector we tested (YOLO11x, trained on the COCO classes) has no class for either. Riders are part of their vehicle, not separate people.

50 boxes
person23
motorbike10
rickshaw7
cng_autorickshaw3
bicycle3
car3
pickup1

Top-down drone photo of a Dhaka street with polygons around each rooftop, tree, car, van and motorbike, and the road as one region — instance_polygon
Rooftops, trees and traffic from straight above
Each building, tree and vehicle is its own polygon. The road is one region, cut around the trees that hang over it.

24 polygons
road1
building13
tree4
car3
van1
motorbike2

Drone photo of an expressway interchange with canals, with the road, water and vegetation each shaded as regions — semantic_mask
Road, water and vegetation at an interchange
Region labels for an expressway interchange with canals on both sides. Unlabelled areas are left as background.

9 regions
vegetation4
road1
water4

Who labels like these are for.

Example user stories written for the three samples above. They are not client projects. Each one starts from what a model's first draft produced, because that is where most annotation work begins.

bounding_box

Traffic counts that know a rickshaw

As a traffic analytics team working in Dhaka, I want CNG auto-rickshaws and cycle rickshaws counted as their own classes, so that the vehicle mix is not folded into car and bicycle.

Model draft: The detector we tested (YOLO11x, COCO classes) drew 72 boxes. Its class list has neither vehicle. It called 15 objects car, and three of them are cars. The others include CNG auto-rickshaws, rickshaws and motorbikes. It also found two potted plants.
After correction: 50 boxes in seven classes. Riders are part of their vehicle, and the false detections are gone.
Acceptance criteria for a real project: Every vehicle above an agreed minimum size has one box in the right class, and nobody riding a vehicle is counted as a pedestrian. This sample was not checked against that criterion: a few small or hidden objects are unlabelled.

instance_polygon

Roof by roof, from straight above

As a rooftop solar team surveying a neighbourhood by drone, I want each roof outlined on its own, so that each visible roof region can feed a later assessment of where panels might fit.

Model draft: The detector returned one box, labelled cell phone, around something that is not a phone. It has no class for roofs, so it found no buildings at all.
After correction: 13 building outlines, drafted with a segmentation model (SAM 2.1) and corrected, with overhanging trees cut out of the roofs.
Acceptance criteria for a real project: Neighbouring roofs never share an outline, and trees over a roof are cut out of its region. Outlines show visible roof, not usable area: shading, structure and obstructions need a separate assessment.

semantic_mask

Water and green cover between flights

As a city planning or flood monitoring team, I want road, water and vegetation mapped as regions, so that I can compare canal extent and green cover between survey flights.

Model draft: The detector found one truck. Point-prompted segmentation spread the eastern canal across the carriageway beside it, so that stretch was redrawn manually.
After correction: 9 regions in three classes, with unclassified ground left as background.
Acceptance criteria for a real project: Classes never overlap, and nothing is forced into a class it does not belong to. This sample has not been independently reviewed against that criterion.

Label these photos yourself

What needs labelling, and what is it for?

Tell us the data type, rough volume if you know it, the labels you need and how the data will be used. If the guidelines are still being written, say so. Please don't include the dataset in your first message.

Discuss your project