ESP_027750_2640_RED Computer Vision Project

Background

Images contained in this project were obtained by segmenting a large sattelite image of the Martian North Polar region. The area captured by the camera is located just beneath the steep slopes of a huge ice mass known as North Polar Layered Deposits (i.e. imagine a glacier that sists right on top of the planet). Avalanches are a regular occurence in the region. Icy boulders roll down the slopes of NPLD and land right into the thick layer of dust/sand just beneath it - these blocks are what we're asking you to label.

Datasets

Prior to uploading the images all negatives (i.e. images with no objects of interest) were filtered out. The remaing photos all contain at least one instance of the boulder class and are organised in the following way:

• weak-positive: contains a few small & sparsely distributed objects
• complex: contains piles of debris that are hard to annotate
• well-defined: contains moderate amounts of well defined objects of different sizes
• demo: a selection of 20 pre-annotated images representing some of the more complex backgrounds one might encounter

Labelling guidlines

The are only 2 lables:

• boulder: blocks of CO2 ice that broke off from NPLD
• cluster: groups of blocks separated by <2 pixels that are hard to differentiate.

When to label:

• Objects that are > 6x6 pixels
• Objects that are lighter than the background and/or produce a clearly visible shadow
• Objects that are not part of some underlying geological structure
• Objects on the edges of the frame that are more than 50% visible
• Groups of objects that are hard to differentiate should be labeled as "clusters".

When NOT to label:

• Objects that are < 6x6 pixels
• Objects that cast no shadow (i.e. white dots/areas)
• Objects that are part of some larger formation (e.g. an elonagted ice ridge as seen in demo 11,18 & 19)
• Objects on the edges of the frame that are less than 50% visible

NOTE: This task could feel very subjective at times, particularly when deciding whether to label small blocks partially covered with sand. A good practice is to zoom out and unfocus your eyes, the objects that stand out against the background should be labelled. It is perfectly fine to miss some of the smaller objects - these will be labeled separately after the images are upscaled.

                        @misc{
                            esp_027750_2640_red_dataset,
                            title = { ESP_027750_2640_RED Dataset },
                            type = { Open Source Dataset },
                            author = { HiRISE },
                            howpublished = { \url{ https://universe.roboflow.com/hirise/esp_027750_2640_red } },
                            url = { https://universe.roboflow.com/hirise/esp_027750_2640_red },
                            journal = { Roboflow Universe },
                            publisher = { Roboflow },
                            year = { 2023 },
                            month = { jan },
                            note = { visited on 2024-11-28 },
                            }