I found that:
\nTrain : horizontal images only
\nTest : horizontal and -90° rotated images
Any thoughts, negative examples?
", "messages": [ { "id": 1231322, "postDate": "2021-03-08T21:53:34.160Z", "content": "I found that:
\nTrain : horizontal images only
\nTest : horizontal and -90° rotated images
Any thoughts, negative examples?
", "rawMarkdown": "I found that:\nTrain : horizontal images only\nTest : horizontal and -90° rotated images\n\n1. Am I right?\n2. Are we supposed to be robust on the -90° rotation?\nIf so, I did not find an image that violates that:\n- if it is wider -> horizontal\n- if it is taller -> rotated (and so just rotate it back 90°)\n\nAny thoughts, negative examples?", "votes": 9 }, { "id": 1231343, "postDate": "2021-03-08T22:09:36.513Z", "content": "I already found a counter-example 888a5efe68d6 243x234 width by height, and despite of being wider than its height, it is rotated.
\nI suppose it looked like that because molecules are mostly drawn so that they are taking up space horizontally so that they can be put in books more easily for example?
I was able to tell that they are rotated 90° and not -90° based on letters B and F. But are all of them so?
\nWhat are your findings on the test set?
If the rotation can not only be 90° but also -90°, then we cannot tell the orientation of compounds that only contain atoms with point symmetric letters.. and so there are two possible compounds (which are the mirrored copies of each other) in the image that we cannot tell apart and I suppose that mostly the two has different formulae.
", "rawMarkdown": "If the rotation can not only be 90° but also -90°, then we cannot tell the orientation of compounds that only contain atoms with point symmetric letters.. and so there are two possible compounds (which are the mirrored copies of each other) in the image that we cannot tell apart and I suppose that mostly the two has different formulae." }, { "id": 1231365, "postDate": "2021-03-08T22:30:12.480Z", "content": "It seems to me that the test set only uses one clockwise rotation, never the CCW direction. I was surprised to see you find a counter-example in the test set. According to my script, all the training images have bigger width than height.
", "rawMarkdown": "It seems to me that the test set only uses one clockwise rotation, never the CCW direction. I was surprised to see you find a counter-example in the test set. According to my script, all the training images have bigger width than height." }, { "id": 1231368, "postDate": "2021-03-08T22:40:50.697Z", "content": "Yeah, it seems that counter-examples are rare… But still, should keep in mind that not 100% true.
\nTrain a network that tells if a picture is rotated, for example. Which is simple to do with the training data. If that finds a lot of counter-examples in the test set, then it might be worth the extra effort to work on this.
\nI was just randomly scoping the training data and found a369aef8699e.
\nIt seems that the training data also has these very few weird rotations.
I already found a counter-example 888a5efe68d6 243x234 width by height, and despite of being wider than its height, it is rotated.
\nI suppose it looked like that because molecules are mostly drawn so that they are taking up space horizontally so that they can be put in books more easily for example?
I was able to tell that they are rotated 90° and not -90° based on letters B and F. But are all of them so?
\nWhat are your findings on the test set?
If the rotation can not only be 90° but also -90°, then we cannot tell the orientation of compounds that only contain atoms with point symmetric letters.. and so there are two possible compounds (which are the mirrored copies of each other) in the image that we cannot tell apart and I suppose that mostly the two has different formulae.
", "votes": 0, "replies": [] }, { "id": 1231365, "author_name": "Matt", "author_url": "", "post_date": "2021-03-08T22:30:12.480000", "content": "It seems to me that the test set only uses one clockwise rotation, never the CCW direction. I was surprised to see you find a counter-example in the test set. According to my script, all the training images have bigger width than height.
", "votes": 0, "replies": [] }, { "id": 1231368, "author_name": "nofreewill42", "author_url": "", "post_date": "2021-03-08T22:40:50.697000", "content": "Yeah, it seems that counter-examples are rare… But still, should keep in mind that not 100% true.
\nTrain a network that tells if a picture is rotated, for example. Which is simple to do with the training data. If that finds a lot of counter-examples in the test set, then it might be worth the extra effort to work on this.
\nI was just randomly scoping the training data and found a369aef8699e.
\nIt seems that the training data also has these very few weird rotations.