I have little to know idea what to expect as far as shakeup goes in this competition and thought I would get some thoughts from you all?
\nHere is my thoughts. This comp is pretty difficult to realistically “train” your own model, however we can validate other models on our dataset and find what works best. By now we have determined that the challenges mentioned in our data card are highly prevalent in the data as suggested by the baseline models still being at the top of the LB a month later. The real question is how much a change in the distribution of challenges can change the LB. With how close all our scores are at this point I think unless you have something special like those at say the 0.16+ mark that there is decent potential for drift? This is just my assumption but I am curious what you all think! And of course it’s still very early but for most of us I don’t see the Lb changing much.
", "messages": [ { "id": "2770319", "postDate": "04/23/2024 19:20:26", "content": "I have little to know idea what to expect as far as shakeup goes in this competition and thought I would get some thoughts from you all?
\nHere is my thoughts. This comp is pretty difficult to realistically “train” your own model, however we can validate other models on our dataset and find what works best. By now we have determined that the challenges mentioned in our data card are highly prevalent in the data as suggested by the baseline models still being at the top of the LB a month later. The real question is how much a change in the distribution of challenges can change the LB. With how close all our scores are at this point I think unless you have something special like those at say the 0.16+ mark that there is decent potential for drift? This is just my assumption but I am curious what you all think! And of course it’s still very early but for most of us I don’t see the Lb changing much.
", "rawMarkdown": "I have little to know idea what to expect as far as shakeup goes in this competition and thought I would get some thoughts from you all? \n\nHere is my thoughts. This comp is pretty difficult to realistically “train” your own model, however we can validate other models on our dataset and find what works best. By now we have determined that the challenges mentioned in our data card are highly prevalent in the data as suggested by the baseline models still being at the top of the LB a month later. The real question is how much a change in the distribution of challenges can change the LB. With how close all our scores are at this point I think unless you have something special like those at say the 0.16+ mark that there is decent potential for drift? This is just my assumption but I am curious what you all think! And of course it’s still very early but for most of us I don’t see the Lb changing much.", "votes": null }, { "id": "2813894", "postDate": "05/15/2024 04:43:21", "content": "I recently participated in this competition, and I think a shakeup is likely to occur due to the congestion of LB and the large proportion of private LB. Also, it seems like the same baseline model is being used.
", "rawMarkdown": "I recently participated in this competition, and I think a shakeup is likely to occur due to the congestion of LB and the large proportion of private LB. Also, it seems like the same baseline model is being used.", "votes": null }, { "id": "2822576", "postDate": "05/18/2024 17:10:21", "content": "The dataset is really overkilled realistic. To gain 0.25+, I think its important to rationalize the objects of not as grid of pixels or features detected by CNNs but as continuous-constructible-generatable objects , while accounting with features such as orientation, day or night, atmospheric refraction, humans and animals. As you point about using pretrained models. They actually seem to be a good idea because the images of transparent glasses, not even help me as a human to 3d reconstruct it in my mind, let alone be an algorithm. It would be far better to detect it as a glass cup and use transfer learning to get a 3D reconstruction. I detected a feature to that I do not think other would ever be using that is to take the path of the image in the hidden test and if the path is labeled that take it and already know what type of thing your model will be trying to 3D construct then grasping a pretrained network it help it out .
", "rawMarkdown": "The dataset is really overkilled realistic. To gain 0.25+, I think its important to rationalize the objects of not as grid of pixels or features detected by CNNs but as continuous-constructible-generatable objects , while accounting with features such as orientation, day or night, atmospheric refraction, humans and animals. As you point about using pretrained models. They actually seem to be a good idea because the images of transparent glasses, not even help me as a human to 3d reconstruct it in my mind, let alone be an algorithm. It would be far better to detect it as a glass cup and use transfer learning to get a 3D reconstruction. I detected a feature to that I do not think other would ever be using that is to take the path of the image in the hidden test and if the path is labeled that take it and already know what type of thing your model will be trying to 3D construct then grasping a pretrained network it help it out .", "votes": null }, { "id": "2830235", "postDate": "05/23/2024 04:57:23", "content": "you are right , it seems like the same baseline model is being used and most of us just adjust parameters
", "rawMarkdown": "you are right , it seems like the same baseline model is being used and most of us just adjust parameters", "votes": null } ], "comments": [ { "id": 2813894, "author_name": "mitsuyasuhoshino", "author_url": "", "post_date": "05/15/2024 04:43:21", "content": "I recently participated in this competition, and I think a shakeup is likely to occur due to the congestion of LB and the large proportion of private LB. Also, it seems like the same baseline model is being used.
", "votes": null, "replies": [ { "id": 2830235, "author_name": "cokkies", "author_url": "", "post_date": "05/23/2024 04:57:23", "content": "you are right , it seems like the same baseline model is being used and most of us just adjust parameters
", "votes": null, "replies": [] } ] }, { "id": 2822576, "author_name": "devsya", "author_url": "", "post_date": "05/18/2024 17:10:21", "content": "The dataset is really overkilled realistic. To gain 0.25+, I think its important to rationalize the objects of not as grid of pixels or features detected by CNNs but as continuous-constructible-generatable objects , while accounting with features such as orientation, day or night, atmospheric refraction, humans and animals. As you point about using pretrained models. They actually seem to be a good idea because the images of transparent glasses, not even help me as a human to 3d reconstruct it in my mind, let alone be an algorithm. It would be far better to detect it as a glass cup and use transfer learning to get a 3D reconstruction. I detected a feature to that I do not think other would ever be using that is to take the path of the image in the hidden test and if the path is labeled that take it and already know what type of thing your model will be trying to 3D construct then grasping a pretrained network it help it out .
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2770319": "I have little to know idea what to expect as far as shakeup goes in this competition and thought I would get some thoughts from you all? \n\nHere is my thoughts. This comp is pretty difficult to realistically “train” your own model, however we can validate other models on our dataset and find what works best. By now we have determined that the challenges mentioned in our data card are highly prevalent in the data as suggested by the baseline models still being at the top of the LB a month later. The real question is how much a change in the distribution of challenges can change the LB. With how close all our scores are at this point I think unless you have something special like those at say the 0.16+ mark that there is decent potential for drift? This is just my assumption but I am curious what you all think! And of course it’s still very early but for most of us I don’t see the Lb changing much.", "2813894": "I recently participated in this competition, and I think a shakeup is likely to occur due to the congestion of LB and the large proportion of private LB. Also, it seems like the same baseline model is being used.", "2822576": "The dataset is really overkilled realistic. To gain 0.25+, I think its important to rationalize the objects of not as grid of pixels or features detected by CNNs but as continuous-constructible-generatable objects , while accounting with features such as orientation, day or night, atmospheric refraction, humans and animals. As you point about using pretrained models. They actually seem to be a good idea because the images of transparent glasses, not even help me as a human to 3d reconstruct it in my mind, let alone be an algorithm. It would be far better to detect it as a glass cup and use transfer learning to get a 3D reconstruction. I detected a feature to that I do not think other would ever be using that is to take the path of the image in the hidden test and if the path is labeled that take it and already know what type of thing your model will be trying to 3D construct then grasping a pretrained network it help it out .", "2830235": "you are right , it seems like the same baseline model is being used and most of us just adjust parameters" }, "source": "meta" }