![\"conceptual](\"https://raw.githubusercontent.com/toshi-k/kaggle-3d-object-detection-for-autonomous-vehicles/master/img/concept.png\")
Hi Everyone,
\n\nEnding final validation, I share my solution.\nI converted points clouds into RGB images and applied simple SSD (Single Shot Multibox Detector).\nThe Brightness corresponds to the density of clouds, and the Hue corresponds to the weighted center of z-position.
\n\nCode: https://github.com/toshi-k/kaggle-3d-object-detection-for-autonomous-vehicles
\n\n
Hi Everyone,
\n\nEnding final validation, I share my solution.\nI converted points clouds into RGB images and applied simple SSD (Single Shot Multibox Detector).\nThe Brightness corresponds to the density of clouds, and the Hue corresponds to the weighted center of z-position.
\n\nCode: https://github.com/toshi-k/kaggle-3d-object-detection-for-autonomous-vehicles
\n\n
Ah, these are RGB images in the BEV domain? cool idea using color for height. I feel PointPillars is effective in part because it maintains some of the x-y shape variation within each pillar. What quantization did you apply for the images? I wonder if this approach would have done slightly better with a very fine quantization or a PointPillar-like encoding of x-y data e.g. in the saturation and/or luminance channels. A bush and a car can look fairly similar unless you have enough resolution to distinguish the wheel wells and such versus the random-ish shape of a bush.
", "rawMarkdown": "Ah, these are RGB images in the BEV domain? cool idea using color for height. I feel PointPillars is effective in part because it maintains some of the x-y shape variation within each pillar. What quantization did you apply for the images? I wonder if this approach would have done slightly better with a very fine quantization or a PointPillar-like encoding of x-y data e.g. in the saturation and/or luminance channels. A bush and a car can look fairly similar unless you have enough resolution to distinguish the wheel wells and such versus the random-ish shape of a bush.", "replies": [ { "id": 684672, "postDate": "2019-11-30T08:21:28.323Z", "content": "Thank you for your comment ! I just counted number of points in each grid for quantization. Althongh It might not be efficient, it is easy to tune hyperparameters because of its simplicity. I had simmilar ideas and tried PointPillars. However I could not tune it enough. I gave up to use it finally.
", "rawMarkdown": "Thank you for your comment ! I just counted number of points in each grid for quantization. Althongh It might not be efficient, it is easy to tune hyperparameters because of its simplicity. I had simmilar ideas and tried PointPillars. However I could not tune it enough. I gave up to use it finally.\n" } ] }, { "id": 683141, "postDate": "2019-11-28T05:54:56.327Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 684416, "author_name": "oarph", "author_url": "", "post_date": "2019-11-29T17:54:04.827000", "content": "Ah, these are RGB images in the BEV domain? cool idea using color for height. I feel PointPillars is effective in part because it maintains some of the x-y shape variation within each pillar. What quantization did you apply for the images? I wonder if this approach would have done slightly better with a very fine quantization or a PointPillar-like encoding of x-y data e.g. in the saturation and/or luminance channels. A bush and a car can look fairly similar unless you have enough resolution to distinguish the wheel wells and such versus the random-ish shape of a bush.
", "votes": 0, "replies": [ { "id": 684672, "author_name": "toshi_k", "author_url": "", "post_date": "2019-11-30T08:21:28.323000", "content": "Thank you for your comment ! I just counted number of points in each grid for quantization. Althongh It might not be efficient, it is easy to tune hyperparameters because of its simplicity. I had simmilar ideas and tried PointPillars. However I could not tune it enough. I gave up to use it finally.
", "votes": 0, "replies": [] } ] }, { "id": 683141, "author_name": "", "author_url": "", "post_date": "2019-11-28T05:54:56.327000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "682384": "Hi Everyone,\n\nEnding final validation, I share my solution.\nI converted points clouds into RGB images and applied simple SSD (Single Shot Multibox Detector).\nThe Brightness corresponds to the density of clouds, and the Hue corresponds to the weighted center of z-position.\n\nCode: https://github.com/toshi-k/kaggle-3d-object-detection-for-autonomous-vehicles\n\n![conceptual diagram][1]\n\n [1]: https://raw.githubusercontent.com/toshi-k/kaggle-3d-object-detection-for-autonomous-vehicles/master/img/concept.png", "684416": "Ah, these are RGB images in the BEV domain? cool idea using color for height. I feel PointPillars is effective in part because it maintains some of the x-y shape variation within each pillar. What quantization did you apply for the images? I wonder if this approach would have done slightly better with a very fine quantization or a PointPillar-like encoding of x-y data e.g. in the saturation and/or luminance channels. A bush and a car can look fairly similar unless you have enough resolution to distinguish the wheel wells and such versus the random-ish shape of a bush.", "683141": "" } }