After augmentations with rotation and brightness I got a worse result that the precision is rather lower than before. Moreover, the AP and precision curves fluctuated dastically. Did I augmentate them too much?
", "messages": [ { "id": "916989", "postDate": "07/06/2020 06:57:19", "content": "After augmentations with rotation and brightness I got a worse result that the precision is rather lower than before. Moreover, the AP and precision curves fluctuated dastically. Did I augmentate them too much?
", "rawMarkdown": "After augmentations with rotation and brightness I got a worse result that the precision is rather lower than before. Moreover, the AP and precision curves fluctuated dastically. Did I augmentate them too much?", "votes": null }, { "id": "916999", "postDate": "07/06/2020 07:01:20", "content": "How many epochs did you do before and after using the augs? \nI usually do a few more epochs when introducing new augs.
", "rawMarkdown": "How many epochs did you do before and after using the augs? \nI usually do a few more epochs when introducing new augs.", "votes": null }, { "id": "917001", "postDate": "07/06/2020 07:03:48", "content": "You could have augmented \"too much\" (making the scene too bright or too dark), a good way to judge that is just to look at the images after augmentation. I'm not too familiar with Yolov5 but doesn't it have all those augmentations built-in to it? If so, adding more augmentations on top would likely not improve your performance (and may decrease it like you see)
", "rawMarkdown": "You could have augmented \"too much\" (making the scene too bright or too dark), a good way to judge that is just to look at the images after augmentation. I'm not too familiar with Yolov5 but doesn't it have all those augmentations built-in to it? If so, adding more augmentations on top would likely not improve your performance (and may decrease it like you see)", "votes": null }, { "id": "917248", "postDate": "07/06/2020 11:11:40", "content": "I modified the original yoloV5, adding some augmentations in it such as random brightness. Before augmentations I tested the effect of augmentations as well. In my mind it had been supposed to be accepted as similiar as the original datasets. Probably I have to omit my custom augmentations :(
", "rawMarkdown": "I modified the original yoloV5, adding some augmentations in it such as random brightness. Before augmentations I tested the effect of augmentations as well. In my mind it had been supposed to be accepted as similiar as the original datasets. Probably I have to omit my custom augmentations :(", "votes": null }, { "id": "917251", "postDate": "07/06/2020 11:12:51", "content": "I trained 100 epochs but it seemed not to converge and than I trained another 50 epochs. But there is not so much improvement.
", "rawMarkdown": "I trained 100 epochs but it seemed not to converge and than I trained another 50 epochs. But there is not so much improvement.", "votes": null } ], "comments": [ { "id": 916999, "author_name": "mightyrains", "author_url": "", "post_date": "07/06/2020 07:01:20", "content": "How many epochs did you do before and after using the augs? \nI usually do a few more epochs when introducing new augs.
", "votes": null, "replies": [ { "id": 917251, "author_name": "jianangu", "author_url": "", "post_date": "07/06/2020 11:12:51", "content": "I trained 100 epochs but it seemed not to converge and than I trained another 50 epochs. But there is not so much improvement.
", "votes": null, "replies": [] } ] }, { "id": 917001, "author_name": "matthewmasters", "author_url": "", "post_date": "07/06/2020 07:03:48", "content": "You could have augmented \"too much\" (making the scene too bright or too dark), a good way to judge that is just to look at the images after augmentation. I'm not too familiar with Yolov5 but doesn't it have all those augmentations built-in to it? If so, adding more augmentations on top would likely not improve your performance (and may decrease it like you see)
", "votes": null, "replies": [ { "id": 917248, "author_name": "jianangu", "author_url": "", "post_date": "07/06/2020 11:11:40", "content": "I modified the original yoloV5, adding some augmentations in it such as random brightness. Before augmentations I tested the effect of augmentations as well. In my mind it had been supposed to be accepted as similiar as the original datasets. Probably I have to omit my custom augmentations :(
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "916989": "After augmentations with rotation and brightness I got a worse result that the precision is rather lower than before. Moreover, the AP and precision curves fluctuated dastically. Did I augmentate them too much?", "916999": "How many epochs did you do before and after using the augs? \nI usually do a few more epochs when introducing new augs.", "917001": "You could have augmented \"too much\" (making the scene too bright or too dark), a good way to judge that is just to look at the images after augmentation. I'm not too familiar with Yolov5 but doesn't it have all those augmentations built-in to it? If so, adding more augmentations on top would likely not improve your performance (and may decrease it like you see)", "917248": "I modified the original yoloV5, adding some augmentations in it such as random brightness. Before augmentations I tested the effect of augmentations as well. In my mind it had been supposed to be accepted as similiar as the original datasets. Probably I have to omit my custom augmentations :(", "917251": "I trained 100 epochs but it seemed not to converge and than I trained another 50 epochs. But there is not so much improvement." }, "source": "meta" }