\nSo take a look at this
\n1) 2 frames behind (detected by the model but not in GT's)
\n
![\"missed\"](\"https://i.ibb.co/DrjZzbd/missed.png\")
\n2) About 6 frames ahead ( detected by the model and in the GT's now )
\n
![\"found\"](\"https://i.ibb.co/9H2vhfv/found.png\")
validate javascript syntax online
Hey guys , I am a pure newbie so I might be wrong, But I found something to be a bit confusing ,
\nSo take a look at this
\n1) 2 frames behind (detected by the model but not in GT's)
\n
\n2) About 6 frames ahead ( detected by the model and in the GT's now )
\n
validate javascript syntax online
There are about 100+ instances just like this , which causes the overall FP to go up even though the detection itself is correct
", "messages": [ { "id": "1660013", "postDate": "01/22/2022 10:11:29", "content": "Hey guys , I am a pure newbie so I might be wrong, But I found something to be a bit confusing ,
\nSo take a look at this
\n1) 2 frames behind (detected by the model but not in GT's)
\n
\n2) About 6 frames ahead ( detected by the model and in the GT's now )
\n
validate javascript syntax online
There are about 100+ instances just like this , which causes the overall FP to go up even though the detection itself is correct
", "rawMarkdown": "Hey guys , I am a pure newbie so I might be wrong, But I found something to be a bit confusing , \nSo take a look at this \n1) 2 frames behind (detected by the model but not in GT's)\n\n2) About 6 frames ahead ( detected by the model and in the GT's now ) \nThe perspective view is changing. It is hard to say when it starts to be a starfish. It would be harmful if starfish is annotated only for a few pixel visibility.
\nThere was some discussion about that. You can find some useful artifacts here: Generating additional labeling data, Missing Annotations
That somewhat true , but again this causes problems training , meanwhile there are some annotations that mark starfish that even I myself can't see
", "rawMarkdown": "That somewhat true , but again this causes problems training , meanwhile there are some annotations that mark starfish that even I myself can't see", "votes": null }, { "id": "1660578", "postDate": "01/22/2022 19:08:03", "content": "I am dead sure that annotators missed some COTS even if they are fully visible. I guess you can adjust your training bboxes (add missed ones). It would increase your model KPI. But, unfortunately, we can't validate the test set.
\nA different story happens about the perspective view of the COTS. Or in other words - when should we start to annotate COTS? After one pixel began visible? Half of COTS? Actually, such an issue is a part of philosophy - imagine two pictures: one of a car, and the second of a horse. Let's do alpha blending from the horse picture to the car. At what alpha value does the horse stop being a horse? We have no answer, it is very subjective…
", "rawMarkdown": "I am dead sure that annotators missed some COTS even if they are fully visible. I guess you can adjust your training bboxes (add missed ones). It would increase your model KPI. But, unfortunately, we can't validate the test set.\n\nA different story happens about the perspective view of the COTS. Or in other words - when should we start to annotate COTS? After one pixel began visible? Half of COTS? Actually, such an issue is a part of philosophy - imagine two pictures: one of a car, and the second of a horse. Let's do alpha blending from the horse picture to the car. At what alpha value does the horse stop being a horse? We have no answer, it is very subjective...", "votes": null } ], "comments": [ { "id": 1660095, "author_name": "meowmeowmeowmeowmeow", "author_url": "", "post_date": "01/22/2022 11:49:56", "content": "The perspective view is changing. It is hard to say when it starts to be a starfish. It would be harmful if starfish is annotated only for a few pixel visibility.
\nThere was some discussion about that. You can find some useful artifacts here: Generating additional labeling data, Missing Annotations
That somewhat true , but again this causes problems training , meanwhile there are some annotations that mark starfish that even I myself can't see
", "votes": null, "replies": [] }, { "id": 1660578, "author_name": "meowmeowmeowmeowmeow", "author_url": "", "post_date": "01/22/2022 19:08:03", "content": "I am dead sure that annotators missed some COTS even if they are fully visible. I guess you can adjust your training bboxes (add missed ones). It would increase your model KPI. But, unfortunately, we can't validate the test set.
\nA different story happens about the perspective view of the COTS. Or in other words - when should we start to annotate COTS? After one pixel began visible? Half of COTS? Actually, such an issue is a part of philosophy - imagine two pictures: one of a car, and the second of a horse. Let's do alpha blending from the horse picture to the car. At what alpha value does the horse stop being a horse? We have no answer, it is very subjective…
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1660013": "Hey guys , I am a pure newbie so I might be wrong, But I found something to be a bit confusing , \nSo take a look at this \n1) 2 frames behind (detected by the model but not in GT's)\n\n2) About 6 frames ahead ( detected by the model and in the GT's now ) \n