Masking Guide Updated PDF

Masking Guide. Since 2020 We capture retail today. So you can see the opportunity of tomorrow 1 © 2017 Trax Image Recognition.All Rights Reserved ❑ What is Masking? Drawing rectangle or squares to cover the...

Masking Guide. Since 2020 We capture retail today. So you can see the opportunity of tomorrow 1 © 2017 Trax Image Recognition.All Rights Reserved ❑ What is Masking? Drawing rectangle or squares to cover the full products, along the border in order to identify separate products by the system. We capture retail today. So you can see the opportunity of tomorrow 2 © 2017 Trax Image Recognition.All Rights Reserved ❑ Why Masking? The goal is to use machine-learning methods to automatically detect label products in a given probe. In Voting, the engine uses rectangles to identify products. To teach the engine we try to use images that are well cropped around a single product. Since we still don't have a magic wand (automatic masking algorithm) to get this done, the service called Masking was created. Masking tasks are generated daily to SMART, QATs then manually draw rectangles (masks) around products as well as they can. Daily Masking Done New voting Masking Voting done Voting in SMART process sent in SMART Evaluation Voting Done request sent We capture retail today. So you can see the opportunity of tomorrow 3 © 2017 Trax Image Recognition.All Rights Reserved ❑ Why Masking? During engine training, after the engine identified the product in a given frame, it has to compare it to a validated label in order to learn for future classifications. Before a probe is sent to teach the engine, a QAT labels all the products in it. There are many products where even the QATs fail to label correctly. That is why voting is done among several people and majority is taken into account. Even then, sometimes it will take more than 7 voters until a majority with high enough confidence achieved. Again, sending all products to many QATs to insure confidence when most products are not as problematic is a waste of resources. To solve this, the voting controller was designed in way that will start each voting process from a low amount of QATs and increases until desired confidence is reached. This is how voting looks when done and printed(each color represent a different QAT): We capture retail today. So you can see the opportunity of tomorrow 4 © 2017 Trax Image Recognition.All Rights Reserved ❑ Why is proper masking important? Even though the tagging is correct, improper masking is a reason for invalid votes in voting. Therefore, proper masking is so important to success this process. The result of bad masking are: Harming engine’s training & accuracy Damaging accuracy of voting PPA Calculation is wrong Hold back pending development such as automatic engine masking Re-masking - A feature that enables sending probes to get re-masking with out losing current labeled data. Key feature for when masking is not done properly. We capture retail today. So you can see the opportunity of tomorrow 5 © 2017 Trax Image Recognition.All Rights Reserved ❑ Why is proper masking so important? Invalid Vote Validation Masking Error Error Not tagging on Multiple Tags Mask is beyond the Mask does not Center of SKU on same SKU SKU cover the SKU Mask intersect the tag We capture retail today. So you can see the opportunity of tomorrow 6 © 2017 Trax Image Recognition.All Rights Reserved Masking Features We capture retail today. So you can see the opportunity of tomorrow 7 © 2017 Trax Image Recognition.All Rights Reserved ❑ Masking Environment Masking will be done in the special environment “Masking” in SMART. We capture retail today. So you can see the opportunity of tomorrow 8 © 2017 Trax Image Recognition.All Rights Reserved ❑ Masking Environment Fills masks with dark background Resize masks Drag masks Change mask Colour We capture retail today. So you can see the opportunity of tomorrow 9 © 2017 Trax Image Recognition.All Rights Reserved ❑ Masking Features - Duplicate Mask Button Masking Features Resize masks By clicking on this icon Change mask Colour you will get duplicate masks We capture retail today. So you can see the opportunity of tomorrow 10 © 2017 Trax Image Recognition.All Rights Reserved ❑ Resize Button SMART Masking - Selecting a mask by double clicking enables the drawing of overlapping masks. By double clicking on the mask will enable the resize buttons, delete button etc. like this We capture retail today. So you can see the opportunity of tomorrow 11 © 2017 Trax Image Recognition.All Rights Reserved ❑ Mask size will be re-adjusted when changing the browser size. When you are changing the browser size masks size will adjust according to the browser size We capture retail today. So you can see the opportunity of tomorrow 12 © 2017 Trax Image Recognition.All Rights Reserved Masking Guidelines We capture retail today. So you can see the opportunity of tomorrow 13 © 2017 Trax Image Recognition.All Rights Reserved ❑ Generic Guidelines All visible products should be masked in the probe considering the generic tagging rules. Mask should be drawn exactly on the SKU boarder. Not too narrow or not too wide Mask on Empty spaces as well. We capture retail today. So you can see the opportunity of tomorrow 14 © 2017 Trax Image Recognition.All Rights Reserved ❑ Should mask all the visible products in the probe considering the tagging rules Maskers should always mask all the visible products in the probe considering the tagging rules. Every SKU which is supposed to be tagged should be masked , if not avoid. Eg: SKU with Red Star are not supposed to be tagged since we need to recognize only the front row. Therefore mask is not required We capture retail today. So you can see the opportunity of tomorrow 15 © 2017 Trax Image Recognition.All Rights Reserved ❑ Margin Make sure to include the margin within the masks (See the correct way in yellow) We capture retail today. So you can see the opportunity of tomorrow 16 © 2017 Trax Image Recognition.All Rights Reserved ❑ Overlap If in case of a products slightly covering up each other - its preferable to draw the masks with a small overlap than drawing a mask smaller as below than the actual product size This way is not preferable We capture retail today. So you can see the opportunity of tomorrow 17 © 2017 Trax Image Recognition.All Rights Reserved ❑ Clip Strip & Hanging Products In this case we have draw mask for each SKU We capture retail today. So you can see the opportunity of tomorrow 18 © 2017 Trax Image Recognition.All Rights Reserved  Click Here for the special masking guideline which ❑ Hoppers/Baskets/Packages available for RWprojects regarding fresh products in One single mask for the entire Hopper/Basket/Package. baskets/boxes & entire shelf without baskets Entire basket or package should be masked. Remember it is Masks should be drawn the same as basket/ Packages. Please better to draw the masks with a little overlap to include the remember that it's better to draw the masks with a little edges. overlap to include the edges. We capture retail today. So you can see the opportunity of tomorrow 19 © 2017 Trax Image Recognition.All Rights Reserved ❑ Transparent packs VS non transparent packs Non-transparent packs should be masked as one unit. Transparent pack should be masked individually Non-Transparent Transparent We capture retail today. So you can see the opportunity of tomorrow 20 © 2017 Trax Image Recognition.All Rights Reserved ❑ Challenges in Transparent Pack Masking In case the products inside the transparent pack are not possible to mask individually due to the small sizes or any other reason, then, drawing a one mask is fine. E.g. We capture retail today. So you can see the opportunity of tomorrow 21 © 2017 Trax Image Recognition.All Rights Reserved ❑ Masking Partial products Follow the tagging rules. If the product to be visible more than its 75% ,You may mask them, otherwise ignore. Do not Mask as the visibility is less than 75% We capture retail today. So you can see the opportunity of tomorrow 22 © 2017 Trax Image Recognition.All Rights Reserved ❑ Masking Reclining Products Mask The reclining products as much as possible considering the reclining products tagging rules. Masking is required only for the reclining products which suppose to be tagged according to the reclining products recognition guide. No need to mask aligning with the reclining recognition rules. We capture retail today. So you can see the opportunity of tomorrow 23 © 2017 Trax Image Recognition.All Rights Reserved ❑ Tilted products Draw one big mask around the group of the tilted products. The rest of the products should be masked normally. We capture retail today. So you can see the opportunity of tomorrow 24 © 2017 Trax Image Recognition.All Rights Reserved  ❑ Masking in rotated probes In case, you have a rotated probe, you should cancel the probe using “Probe is rotated” option Cancel the probe -No need to mask Important Note: This will avoid receiving rotated probes in "Engine Voting" tasks(EV tasks do not have ability to rotate the probe). In case if you receiving rotated probes in EV tasks, you can cancel those tasks using "Probe is Rotated" option. We capture retail today. So you can see the opportunity of tomorrow 25 © 2017 Trax Image Recognition.All Rights Reserved Masking POSM We capture retail today. So you can see the opportunity of tomorrow 26 © 2017 Trax Image Recognition.All Rights Reserved ❑ Masking POSMs All the POSMs should be masked We capture retail today. So you can see the opportunity of tomorrow 27 © 2017 Trax Image Recognition.All Rights Reserved ❑ Masking POSMs O r We capture retail today. So you can see the opportunity of tomorrow © 2017 Trax Image Recognition. All Rights Reserved ❑ Masking POSMs POSM masking should be avoid ,Only if the products are covering by a POSM. You may Mask only the products which are arranged inside the POSM individually. But do not mask the POSM in this case only. Do not Mask the highlighted POSM We capture retail today. So you can see the opportunity of tomorrow 29 © 2017 Trax Image Recognition.All Rights Reserved Masking Probes with Quality Issues We capture retail today. So you can see the opportunity of tomorrow 30 © 2017 Trax Image Recognition.All Rights Reserved ❑ Masking blurry probes When the probe is too blurry to do masking correctly, You can click done without masking. You have to mask the probe if there are some SKUs which are clear, in a blurry probe. No need to mask We capture retail today. So you can see the opportunity of tomorrow 31 © 2017 Trax Image Recognition.All Rights Reserved ❑ Images taken from too far In case of images taken from a big distance where it is impossible to recognize SKUs, QAT can skip the masking and continue to the next probe. No need to mask We capture retail today. So you can see the opportunity of tomorrow 32 © 2017 Trax Image Recognition.All Rights Reserved ❑ Mask only visible SKUs In a probe, QAT needs to mask only visible products. If there are unclear products in the probes, then QAT can ignore those unclear products. We capture retail today. So you can see the opportunity of tomorrow 33 © 2017 Trax Image Recognition.All Rights Reserved Masking Empty Spaces. We capture retail today. So you can see the opportunity of tomorrow 34 © 2017 Trax Image Recognition.All Rights Reserved ❑ Masking Empty QATs should mask all the empty spaces – where possible to put at least one product. The mask should be drawn around the entire empty space on the shelf. We capture retail today. So you can see the opportunity of tomorrow 35 © 2017 Trax Image Recognition.All Rights Reserved ❑ Masking Empty We capture retail today. So you can see the opportunity of tomorrow 36 © 2017 Trax Image Recognition. All Rights Reserved ❑ Not Empty - When to NOT draw mask on empty space Not Empty We capture retail today. So you can see the opportunity of tomorrow 37 © 2017 Trax Image Recognition. All Rights Reserved Masking Price Labels. We capture retail today. So you can see the opportunity of tomorrow 38 © 2017 Trax Image Recognition.All Rights Reserved ❑ Masking Price Labels. This guidelines are relevant only for Highlighted tile. We capture retail today. So you can see the opportunity of tomorrow 39 © 2017 Trax Image Recognition.All Rights Reserved  ❑ Masking Price Labels. Mask Only the price labels which are visible in the probe. You are not required to mask on the Products here. NOTE: Furthermore, we need to consider the special guidelines which might be provided by R&D team, if there is a special requirement. We capture retail today. So you can see the opportunity of tomorrow 40 © 2017 Trax Image Recognition.All Rights Reserved Masking Menus. We capture retail today. So you can see the opportunity of tomorrow 41 © 2017 Trax Image Recognition.All Rights Reserved ❑ Masking Menus. This guidelines are relevant only for Highlighted tile. We capture retail today. So you can see the opportunity of tomorrow 42 © 2017 Trax Image Recognition.All Rights Reserved  ❑ Masking Menus. We need to mask the menu items one by one considering the menu recognition rules. NOTE: Furthermore, we need to consider the special guidelines which might be provided by R&D team, if there is a special requirement. We capture retail today. So you can see the opportunity of tomorrow 43 © 2017 Trax Image Recognition.All Rights Reserved Bad Masking We capture retail today. So you can see the opportunity of tomorrow 44 © 2017 Trax Image Recognition.All Rights Reserved ❑ Products without masks. The most common and the one that cause the biggest problems – products without masks at all. Green/red rectangles are the masks done by the QAT. Yellow area are the ones that has not been masked at all. We capture retail today. So you can see the opportunity of tomorrow 45 © 2017 Trax Image Recognition.All Rights Reserved ❑ Mask Quality. Mask quality – most of the time too narrow: We capture retail today. So you can see the opportunity of tomorrow 46 © 2017 Trax Image Recognition.All Rights Reserved ❑ Mask Quality. Mask quality – most of the time too narrow: We capture retail today. So you can see the opportunity of tomorrow 47 © 2017 Trax Image Recognition.All Rights Reserved ❑ Clicking “Done” without drawing masks at all P&GCN- 6922959 We capture retail today. So you can see the opportunity of tomorrow 48 © 2017 Trax Image Recognition.All Rights Reserved  Partial Masking From SMART V3.42.0 (DIGI-49) We capture retail today. So you can see the opportunity of tomorrow 49 © 2017 Trax Image Recognition.All Rights Reserved ❑ Partial Masking Example: As part of the partial voting epic we would like to support also partial masking tasks which will feed the partial voting. Meaning that we will send masking requests only on products which will require voting and by that improve SLA. SKUs which will require masking will be marked with a tag. Hence, Masking should be done only on the products with tags – rest of the products shouldn’t be masked. If you try to draw a mask around other products which doesn’t have tags, live alert will popped up and that incorrect masking will disappear immediately. We capture retail today. So you can see the opportunity of tomorrow 50 © 2017 Trax Image Recognition.All Rights Reserved  Rectification for masking tasks From SMART V3.56.0 (DIGI-2072) We capture retail today. So you can see the opportunity of tomorrow 51 © 2017 Trax Image Recognition.All Rights Reserved  ❑ Rectification for masking tasks One of the main KPI for Tesco and Aldi is "Fresh" recognition. For this category, we use mainly Dome cameras. The issue with Dome cameras is that they have a difficult angle for horizontal “bounding box” (SMART Masking) which causes many errors and resulting in invalid voting for accuracy measurement. To support it, we have developed a mechanism in SMART that rectifies those images. The rectification is done before we create a masking task, so the maskers will be able to mask the rectified images. After the masking task is done, the voting task has been done on the original probe. when SMART receives a masking task that should be rectified -> we rectify the image the masker will add the masks, and, in the saving, we are un-rectifying the masks and save them in 2 ways: By 2 points for backward compatibility By 4 points as the new mask after un-rectification is not a square You can see a few examples of rectified probes:  Delete masks when screen size is changed From SMART V3.62.0 (DIGI-5822) We capture retail today. So you can see the opportunity of tomorrow 53 © 2017 Trax Image Recognition.All Rights Reserved  ❑ Delete masks when screen size is changed Since the resolution of the window should be kept stable during the Masking process to avoid the mentioned issue(masking offset) which was if the masker changes the screen size during the task. Therefore, we have implemented a solution to the SMART masking process to delete the masks and notify the maskers with a pop-up message if the masker change the screen size during the task. Hence, the masker should be re-draw all masks for the relevant task. Special Masking Guidelines for Mondelez_US_PS We capture retail today. So you can see the opportunity of tomorrow 55 © 2017 Trax Image Recognition.All Rights Reserved  Piles Masking Unclear Piles (bad ordered) should be masked as ONE UNIT. Incorrect Correct  Piles Masking Unclear Piles (bad ordered) should be masked as ONE UNIT. Incorrect Correct  Piles masking Unclear Piles (bad ordered) should be masked as ONE UNIT. Incorrect Correct  Confectionary masking Small Confectionary products laying on top of each other should be MASKED ONCE, as one unit. Incorrect Correct  Confectionary masking Small Confectionary products laying on top of each other should be MASKED ONCE, as one unit. Incorrect Correct  Cookies & Crackers Piles of Cookies and Crackers, well ordered on top of each other, that can be easily differentiated, should be masked SEPARATELY. Incorrect Correct  FRONT PRODUCTS Only the FACING/FRONT products should be masked Incorrect Correct  BOXES Boxes with Multiply small Products in them should be masked as ONE UNIT Incorrect Correct  BOXES Boxes with Multiply small Products in them should be masked as ONE UNIT Incorrect Correct  BOXES Boxes with Multiply small Products in them should be masked as ONE UNIT Incorrect Correct Full Guided Masking (FGM) The flow: The voting flow will change as the follow: 1st cycle of voting will be initiated prior to the masking, Masking task will follow the voting task and if needed, another cycle of voting will be initiated after the masking: The SLA for both Voting and Masking is the same as the previous of 24hours Full Guided Masking (FGM) Process: Since maskers now able to see the products tags, they can mask according to the tagged products(without having any recognition confusion) in the masking task. Full Guided Masking (FGM) – Empty Masking Empty Masking Guidelines: QATs should mask all the empty space separately with one tag each – following the full Guided masking Adoption ID: DIGI-7901 Full Guided Masking (FGM) – Transparent packs Transparent packs should be masked according to the following guidelines since we can now see the validators tags on the probe in masking with FGM-activated projects. Scenario 01. Transparent pack with tags on each product: Masking: The masker should draw masks for each product with a tag. Note: The above could be challenging in the case of small products. (i.e., Toothbrushes ). Full Guided Masking (FGM) – Transparent packs Scenario 02. Transparent pack with one tag on the pack: Masking: The masker should draw one mask for the entire transparent package. Mssing Mask Alert – PM & FGM Alert 01. Missing masks: Not all tags are masked If just certain tags in the partial masking or FGM probe are dislocated/unnecessary or duplicated, choose "Confirm and Continue" only after ensuring that all relevant tags are masked except for the dislocated/unnecessary or duplicated tags. Date: 01-30-2022 Version: SMART V3.78 Reference: DIGI-9674 Mssing Mask Alert – PM & FGM Alert 02. Missing masks: Zero masks were drawn When you have only dislocated/unnecessary tags in the probes which shouldn't be masked, you can cancel the probe by giving the relevant reason: Special note: only need to cancel the probe if there are dislocated or unnecessary tags, and make sure to mention the cancelation reason properly. Date: 01-30-2022 Version: SMART V3.78 Reference: DIGI-9674 Thank you We capture retail today. So you can see the opportunity of tomorrow 73 © 2017 Trax Image Recognition.All Rights Reserved