Since overfitting strategy failed, let's just talk about what I've done one month ago. π
\nI use CenterNet with DeepLabV3+ architecture and EfficientNetV2 backbone, start from this example.
\nActually, I am pleased to survive in huge shake up. π
\nThanks everyone and congratulation to all winners.
", "messages": [ { "id": "1690578", "postDate": "02/15/2022 02:37:35", "content": "Since overfitting strategy failed, let's just talk about what I've done one month ago. π
\nI use CenterNet with DeepLabV3+ architecture and EfficientNetV2 backbone, start from this example.
\nActually, I am pleased to survive in huge shake up. π
\nThanks everyone and congratulation to all winners.
", "rawMarkdown": "Since overfitting strategy failed, let's just talk about what I've done one month ago. π \n\nI use CenterNet with DeepLabV3+ architecture and EfficientNetV2 backbone, start from this [example](https://keras.io/examples/vision/deeplabv3_plus/).\n- change backbone to EfficientNetV2 B0~XL\n- change output heatmap size to 1/8 input, and add regression head\n- training on 1280x720, inference on 1792x1008(1.4x)\n- blend two can get private LB score 0.712, and 0.722 if inference on 1.6x\n\nActually, I am pleased to survive in huge shake up. π\n\nThanks everyone and congratulation to all winners.", "votes": null }, { "id": "1690593", "postDate": "02/15/2022 02:45:18", "content": "Congratulations @outrunner
", "rawMarkdown": "Congratulations @outrunner", "votes": null }, { "id": "1690594", "postDate": "02/15/2022 02:45:27", "content": "Simple description but definitely contains many experiences and thoughts inside! Thank you for sharing and amazing 1st place public LB position!
", "rawMarkdown": "Simple description but definitely contains many experiences and thoughts inside! Thank you for sharing and amazing 1st place public LB position!", "votes": null }, { "id": "1690597", "postDate": "02/15/2022 02:49:06", "content": "thank you π
", "rawMarkdown": "thank you π", "votes": null }, { "id": "1690598", "postDate": "02/15/2022 02:50:49", "content": "I don't like data inconsistency, but you know, it is kaggle style.π
", "rawMarkdown": "I don't like data inconsistency, but you know, it is kaggle style.π ", "votes": null }, { "id": "1690623", "postDate": "02/15/2022 03:14:59", "content": "Congrat for your money medal! Can you share your training code?
", "rawMarkdown": "Congrat for your money medal! Can you share your training code?", "votes": null }, { "id": "1690634", "postDate": "02/15/2022 03:29:41", "content": "Thanks. Nothing special, just heavily augmentation, perspective transform, rotation, flip, mixup, random crop and batch size every step. Focal loss for heatmap, Huber loss for regression, and Adam optimizer.
", "rawMarkdown": "Thanks. Nothing special, just heavily augmentation, perspective transform, rotation, flip, mixup, random crop and batch size every step. Focal loss for heatmap, Huber loss for regression, and Adam optimizer.", "votes": null }, { "id": "1690639", "postDate": "02/15/2022 03:31:52", "content": "a bit confused, CenterNet is for Object Detection and DeepLabV3+ is for Segmentation. So, how did you use both of them?
You survived this shake up/down! I overfitted to public LB, but I got a bronze medal. Congrats on 4th place! Thanks for sharing your solution.
", "rawMarkdown": "You survived this shake up/down! I overfitted to public LB, but I got a bronze medal. Congrats on 4th place! Thanks for sharing your solution.", "votes": null }, { "id": "1690643", "postDate": "02/15/2022 03:36:45", "content": "I don't know whether I should call it centernet, but objects as points model just like segmentation + regression head.
", "rawMarkdown": "I don't know whether I should call it centernet, but objects as points model just like segmentation + regression head.", "votes": null }, { "id": "1690644", "postDate": "02/15/2022 03:38:12", "content": "WoW. How did you get the mask?
", "rawMarkdown": "WoW. How did you get the mask?", "votes": null }, { "id": "1690647", "postDate": "02/15/2022 03:39:46", "content": "Thanks, it's all luck. π
", "rawMarkdown": "Thanks, it's all luck. π ", "votes": null }, { "id": "1690650", "postDate": "02/15/2022 03:41:11", "content": "reference CenterNet paper to gen the mask
", "rawMarkdown": "reference CenterNet paper to gen the mask", "votes": null }, { "id": "1690657", "postDate": "02/15/2022 03:43:01", "content": "I think there is also CenterNetV2. Did you try that also?
", "rawMarkdown": "I think there is also CenterNetV2. Did you try that also?", "votes": null }, { "id": "1690663", "postDate": "02/15/2022 03:46:16", "content": "No, I did not try it.
", "rawMarkdown": "No, I did not try it.", "votes": null }, { "id": "1690664", "postDate": "02/15/2022 03:46:17", "content": "Congratulation on Solo Win
\nCan you talk about your validation strategy?
Thank youπ
", "rawMarkdown": "Thank youπ", "votes": null }, { "id": "1690673", "postDate": "02/15/2022 03:54:16", "content": "as usual I use 1 fold.
\nmy val set have 4 sequences: [26651, 59337, 53708, 45015]
\nuse all cots frames and half none-cots frames for saving time.
\nselect checkpoints by local F2 score and public LB score.
Thanks @outrunner How did you come about selecting those specific sequences
", "rawMarkdown": "Thanks @outrunner How did you come about selecting those specific sequences", "votes": null }, { "id": "1690697", "postDate": "02/15/2022 04:11:52", "content": "do you use Dynamichead or any other recently published detection head?
", "rawMarkdown": "do you use Dynamichead or any other recently published detection head?", "votes": null }, { "id": "1690701", "postDate": "02/15/2022 04:13:52", "content": "random selection according to mean and total cots counts, then decide by my mood. π
", "rawMarkdown": "random selection according to mean and total cots counts, then decide by my mood. π", "votes": null }, { "id": "1690711", "postDate": "02/15/2022 04:20:47", "content": "Congratulations ! :) great work
", "rawMarkdown": "Congratulations ! :) great work", "votes": null }, { "id": "1690721", "postDate": "02/15/2022 04:29:08", "content": "Congratulations. Winning gold medal requires trying new things and make it work. Thanks for sharing your solution!
", "rawMarkdown": "Congratulations. Winning gold medal requires trying new things and make it work. Thanks for sharing your solution!", "votes": null }, { "id": "1690725", "postDate": "02/15/2022 04:30:24", "content": "No, just features -> conv -> conv -> heatmap
", "rawMarkdown": "No, just features -> conv -> conv -> heatmap", "votes": null }, { "id": "1690726", "postDate": "02/15/2022 04:31:01", "content": "Thank you π
", "rawMarkdown": "Thank you π", "votes": null }, { "id": "1690730", "postDate": "02/15/2022 04:34:23", "content": "Thanks, I like objects as points method because it is easy to apply post processing.
", "rawMarkdown": "Thanks, I like objects as points method because it is easy to apply post processing.", "votes": null }, { "id": "1690749", "postDate": "02/15/2022 04:50:05", "content": "Congratulations. thanks for your sharing solution.
\nJust curious, is your solution implemented with TF Object Detection API? or in Keras + Tensorflow?
tf.keras, thanks.
", "rawMarkdown": "tf.keras, thanks.", "votes": null }, { "id": "1690863", "postDate": "02/15/2022 06:14:58", "content": "double winner!!!
", "rawMarkdown": "double winner!!!", "votes": null }, { "id": "1690895", "postDate": "02/15/2022 06:34:27", "content": "although I have model that can get 0.7+ within 45min, but I did not choose it. π
", "rawMarkdown": "although I have model that can get 0.7+ within 45min, but I did not choose it. π ", "votes": null }, { "id": "1690901", "postDate": "02/15/2022 06:40:15", "content": "Congratulations @outrunner Your scores (pub/priv) are outstanding. Everyday we checked LB looking for your improvements and β¦ asking what you did to cross 0.8 β¦ You are TOP Kaggler! Reading your solution description and I am impressed with the path you chose. ππππͺπͺπππ
", "rawMarkdown": "Congratulations @outrunner Your scores (pub/priv) are outstanding. Everyday we checked LB looking for your improvements and ... asking what you did to cross 0.8 ... You are TOP Kaggler! Reading your solution description and I am impressed with the path you chose. ππππͺπͺπππ", "votes": null }, { "id": "1690912", "postDate": "02/15/2022 06:46:11", "content": "@outrunner I gotta learn this mood based decision from you π
\nCongratulations !
haha, thanks. π
", "rawMarkdown": "haha, thanks. π", "votes": null }, { "id": "1690936", "postDate": "02/15/2022 06:58:23", "content": "I have learned a lot from your sharing, thank you.
", "rawMarkdown": "I have learned a lot from your sharing, thank you.", "votes": null }, { "id": "1690941", "postDate": "02/15/2022 06:59:20", "content": "Good job bro!!!
", "rawMarkdown": "Good job bro!!!", "votes": null }, { "id": "1690962", "postDate": "02/15/2022 07:09:08", "content": "Congrats on yet anogther solo gold!
", "rawMarkdown": "Congrats on yet anogther solo gold!", "votes": null }, { "id": "1690976", "postDate": "02/15/2022 07:14:01", "content": "@outrunner Congratulations! Turns out we were doing very similar thing, mine was centernet with unet, 1/4 input, decoupled head and many more tricks. I also tried EffNetV2 as backbone but NF-backbone was better on cv. Did you use separate offset head on 1/8? Did you train on full frames? How long did it take? I did it on crops, even 256x256 with specific sampling was enough.
\nNote to organizers: this setup can be easily ported to TRT to deploy on edge devices. I did it many times with lighter backbones on pretty much every jetson available (excluding Nano, including AGX). At the same time converting usual anchor-based net will be a nightmare.
", "rawMarkdown": "outrunner Congratulations! Turns out we were doing very similar thing, mine was centernet with unet, 1/4 input, decoupled head and many more tricks. I also tried EffNetV2 as backbone but NF-backbone was better on cv. Did you use *separate* offset head on 1/8? Did you train on full frames? How long did it take? I did it on crops, even 256x256 with specific sampling was enough. \n\nNote to organizers: this setup can be easily ported to TRT to deploy on edge devices. I did it many times with lighter backbones on pretty much every jetson available (excluding Nano, including AGX). At the same time converting usual anchor-based net will be a nightmare.", "votes": null }, { "id": "1690989", "postDate": "02/15/2022 07:19:31", "content": "Thank you π
", "rawMarkdown": "Thank you π", "votes": null }, { "id": "1690990", "postDate": "02/15/2022 07:21:50", "content": "@awsaf49 I build second stage on top on centernet arch, not fully probabilistic as in CNV2 paper thou. It helps a little.
", "rawMarkdown": "awsaf49 I build second stage on top on centernet arch, not fully probabilistic as in CNV2 paper thou. It helps a little.", "votes": null }, { "id": "1691023", "postDate": "02/15/2022 07:34:08", "content": "I train on some crop sizes to full frame randomly. For example, [high, width, batch_size] in [[448, 640, 20], [640, 896, 10], [720, 1280, 6]].
\nTraining takes about 1 day on single 2080ti.
Wow thats a long time. Same again, but I never went above 640, 896. What about offset head, i updated question a little
", "rawMarkdown": "Wow thats a long time. Same again, but I never went above 640, 896. What about offset head, i updated question a little", "votes": null }, { "id": "1691046", "postDate": "02/15/2022 07:48:53", "content": "x = Conv2D(256, 3)(features)
\nx = Conv2D(256, 3)(x)
\nreg = Conv2D(4, 1)(x)
\noffset + wh together π
Same, sorry for never-ending questions, As out pipelines are similar, im trying to understand was it bad coding or something else on my part. Did you do just regression on wh, like dirac-delta distribution, or something cooler like gauss, etc?
", "rawMarkdown": "Same, sorry for never-ending questions, As out pipelines are similar, im trying to understand was it bad coding or something else on my part. Did you do just regression on wh, like dirac-delta distribution, or something cooler like gauss, etc?", "votes": null }, { "id": "1691057", "postDate": "02/15/2022 07:57:29", "content": "oh, I see your update. offset head also is 1/8.
\nI post the model here
But, why 2 stage segment task can make such a improvement to detection?
\nI haven't seen this as a trick or methodβ¦.
base = keras_efficientnet_v2.EfficientNetV2B0(input_shape=image_size, drop_connect_rate=0, pretrained=\"imagenet21k\")
\nx = base.get_layer(\"stack_5_block0_sortcut_swish\").output
\nx = DilatedSpatialPyramidPooling(x)
input_a = UpSampling2D(size=(2, 2), interpolation=\"bilinear\")(x)
\ninput_b = base.get_layer('stack_3_block0_sortcut_swish').output
\ninput_b = convolution_block(input_b, num_filters=48, kernel_size=1)
features = Concatenate(axis=-1)([input_a, input_b])
\nx = convolution_block(features)
\nx = convolution_block(x)
\nheatmap = Conv2D(num_classes, 1, padding=\"same\", activation='sigmoid', name='hm', dtype=tf.float32)(x)
x = Conv2D(256, 3, padding=\"same\", activation='swish')(features)
\nx = Conv2D(256, 3, padding=\"same\", activation='swish')(x)
\nreg = Conv2D(4, 1, padding=\"same\", activation='linear', name='re', dtype=tf.float32)(x)
model = keras.Model(inputs=base.input, outputs={'hm':heatmap, 're':reg})
", "rawMarkdown": "base = keras_efficientnet_v2.EfficientNetV2B0(input_shape=image_size, drop_connect_rate=0, pretrained=\"imagenet21k\")\n\nx = base.get_layer(\"stack_5_block0_sortcut_swish\").output\nx = DilatedSpatialPyramidPooling(x)\n\ninput_a = UpSampling2D(size=(2, 2), interpolation=\"bilinear\")(x)\ninput_b = base.get_layer('stack_3_block0_sortcut_swish').output\ninput_b = convolution_block(input_b, num_filters=48, kernel_size=1)\n\nfeatures = Concatenate(axis=-1)([input_a, input_b])\n\nx = convolution_block(features)\nx = convolution_block(x)\nheatmap = Conv2D(num_classes, 1, padding=\"same\", activation='sigmoid', name='hm', dtype=tf.float32)(x)\n\nx = Conv2D(256, 3, padding=\"same\", activation='swish')(features)\nx = Conv2D(256, 3, padding=\"same\", activation='swish')(x)\nreg = Conv2D(4, 1, padding=\"same\", activation='linear', name='re', dtype=tf.float32)(x)\n\nmodel = keras.Model(inputs=base.input, outputs={'hm':heatmap, 're':reg})", "votes": null }, { "id": "1691127", "postDate": "02/15/2022 08:41:13", "content": "Thanks
\nbut what I most wonder is
\nFPN is also has a upsampling and feture concat,
\nI want to know why this methold have such a advantange than FPN ?
To my understand heatmap mask is generate by centernet
\nyou 2 stage model use centernet output and gt bbox as training target for make better use of object segmentation detail information
\nπ in next competition I will use this in my pipline too.
", "rawMarkdown": "Thanks \nbut what I most wonder is \nFPN is also has a upsampling and feture concat,\nI want to know why this methold have such a advantange than FPN ?\n\nTo my understand heatmap mask is generate by centernet\n\nyou 2 stage model use centernet output and gt bbox as training target for make better use of object segmentation detail information\n\nπ in next competition I will use this in my pipline too.", "votes": null }, { "id": "1691130", "postDate": "02/15/2022 08:45:30", "content": "I did not try other method, so I can not confirm whether it has advantage.
", "rawMarkdown": "I did not try other method, so I can not confirm whether it has advantage.", "votes": null }, { "id": "1691280", "postDate": "02/15/2022 10:29:12", "content": "Thanks for the solution and congrats!
", "rawMarkdown": "Thanks for the solution and congrats!", "votes": null }, { "id": "1691287", "postDate": "02/15/2022 10:33:25", "content": "Good job bro keep it up !!
", "rawMarkdown": "Good job bro keep it up !!", "votes": null }, { "id": "1691316", "postDate": "02/15/2022 10:48:48", "content": "Congratulations! May I ask why you went with CenterNet instead of YOLO or fasterRCNN? Did you try multiple models out and found CenterNet gives the best result?
", "rawMarkdown": "Congratulations! May I ask why you went with CenterNet instead of YOLO or fasterRCNN? Did you try multiple models out and found CenterNet gives the best result?", "votes": null }, { "id": "1691530", "postDate": "02/15/2022 13:10:20", "content": "Can you please share cv score on that fold of 4 sequences
", "rawMarkdown": "Can you please share cv score on that fold of 4 sequences", "votes": null }, { "id": "1691935", "postDate": "02/15/2022 17:36:04", "content": "congratulations @outrunner ππ
", "rawMarkdown": "congratulations @outrunner ππ", "votes": null }, { "id": "1691998", "postDate": "02/15/2022 18:24:53", "content": "\n\nmy val set have 4 sequences: [26651, 59337, 53708, 45015]
\n
interesting!
\nour valid set consists of 5 seq., where 3/5 are similar to yours.
\nThis gave correlation with PVT ~90%
\nRef: 11th solution
@bakeryproducts cv score ranged around 0.78~0.8 @ 1.4x to 2.0x inference size. I the last month, I build a wide range solution to face the possible cots size variance in test set, but it is not the case finally.
\n@imeintanis interesting and congratulations π
", "rawMarkdown": "bakeryproducts cv score ranged around 0.78~0.8 @ 1.4x to 2.0x inference size. I the last month, I build a wide range solution to face the possible cots size variance in test set, but it is not the case finally.\n\n@imeintanis interesting and congratulations π", "votes": null }, { "id": "1692339", "postDate": "02/16/2022 01:30:29", "content": "Thank you π
", "rawMarkdown": "Thank you π", "votes": null }, { "id": "1692343", "postDate": "02/16/2022 01:34:34", "content": "No, I think anchor free base will perform well since the target is simple. (shape, texture, size, etc.)
", "rawMarkdown": "No, I think anchor free base will perform well since the target is simple. (shape, texture, size, etc.)", "votes": null }, { "id": "1692356", "postDate": "02/16/2022 01:57:47", "content": "Great Works Keep it up!
", "rawMarkdown": "Great Works Keep it up!", "votes": null }, { "id": "1928051", "postDate": "09/06/2022 07:50:29", "content": "Is the code publicly available?
", "rawMarkdown": "Is the code publicly available?", "votes": null } ], "comments": [ { "id": 1690593, "author_name": "robsonsan", "author_url": "", "post_date": "02/15/2022 02:45:18", "content": "Congratulations @outrunner
", "votes": null, "replies": [ { "id": 1690597, "author_name": "outrunner", "author_url": "", "post_date": "02/15/2022 02:49:06", "content": "thank you π
", "votes": null, "replies": [] } ] }, { "id": 1690594, "author_name": "vincentwang25", "author_url": "", "post_date": "02/15/2022 02:45:27", "content": "Simple description but definitely contains many experiences and thoughts inside! Thank you for sharing and amazing 1st place public LB position!
", "votes": null, "replies": [ { "id": 1690598, "author_name": "outrunner", "author_url": "", "post_date": "02/15/2022 02:50:49", "content": "I don't like data inconsistency, but you know, it is kaggle style.π
", "votes": null, "replies": [] } ] }, { "id": 1690623, "author_name": "magiccard", "author_url": "", "post_date": "02/15/2022 03:14:59", "content": "Congrat for your money medal! Can you share your training code?
", "votes": null, "replies": [ { "id": 1690634, "author_name": "outrunner", "author_url": "", "post_date": "02/15/2022 03:29:41", "content": "Thanks. Nothing special, just heavily augmentation, perspective transform, rotation, flip, mixup, random crop and batch size every step. Focal loss for heatmap, Huber loss for regression, and Adam optimizer.
", "votes": null, "replies": [] }, { "id": 1690666, "author_name": "magiccard", "author_url": "", "post_date": "02/15/2022 03:47:36", "content": "Thank youπ
", "votes": null, "replies": [] } ] }, { "id": 1690639, "author_name": "awsaf49", "author_url": "", "post_date": "02/15/2022 03:31:52", "content": "a bit confused, CenterNet is for Object Detection and DeepLabV3+ is for Segmentation. So, how did you use both of them?
I don't know whether I should call it centernet, but objects as points model just like segmentation + regression head.
", "votes": null, "replies": [] }, { "id": 1690644, "author_name": "awsaf49", "author_url": "", "post_date": "02/15/2022 03:38:12", "content": "WoW. How did you get the mask?
", "votes": null, "replies": [] }, { "id": 1690650, "author_name": "outrunner", "author_url": "", "post_date": "02/15/2022 03:41:11", "content": "reference CenterNet paper to gen the mask
", "votes": null, "replies": [] }, { "id": 1690657, "author_name": "awsaf49", "author_url": "", "post_date": "02/15/2022 03:43:01", "content": "I think there is also CenterNetV2. Did you try that also?
", "votes": null, "replies": [] }, { "id": 1690663, "author_name": "outrunner", "author_url": "", "post_date": "02/15/2022 03:46:16", "content": "No, I did not try it.
", "votes": null, "replies": [] }, { "id": 1690990, "author_name": "bakeryproducts", "author_url": "", "post_date": "02/15/2022 07:21:50", "content": "@awsaf49 I build second stage on top on centernet arch, not fully probabilistic as in CNV2 paper thou. It helps a little.
", "votes": null, "replies": [] } ] }, { "id": 1690642, "author_name": "ttkagglett", "author_url": "", "post_date": "02/15/2022 03:36:45", "content": "You survived this shake up/down! I overfitted to public LB, but I got a bronze medal. Congrats on 4th place! Thanks for sharing your solution.
", "votes": null, "replies": [ { "id": 1690647, "author_name": "outrunner", "author_url": "", "post_date": "02/15/2022 03:39:46", "content": "Thanks, it's all luck. π
", "votes": null, "replies": [] } ] }, { "id": 1690664, "author_name": "morizin", "author_url": "", "post_date": "02/15/2022 03:46:17", "content": "Congratulation on Solo Win
\nCan you talk about your validation strategy?
as usual I use 1 fold.
\nmy val set have 4 sequences: [26651, 59337, 53708, 45015]
\nuse all cots frames and half none-cots frames for saving time.
\nselect checkpoints by local F2 score and public LB score.
Thanks @outrunner How did you come about selecting those specific sequences
", "votes": null, "replies": [] }, { "id": 1690701, "author_name": "outrunner", "author_url": "", "post_date": "02/15/2022 04:13:52", "content": "random selection according to mean and total cots counts, then decide by my mood. π
", "votes": null, "replies": [] }, { "id": 1690912, "author_name": "nyleve", "author_url": "", "post_date": "02/15/2022 06:46:11", "content": "@outrunner I gotta learn this mood based decision from you π
\nCongratulations !
haha, thanks. π
", "votes": null, "replies": [] }, { "id": 1691530, "author_name": "bakeryproducts", "author_url": "", "post_date": "02/15/2022 13:10:20", "content": "Can you please share cv score on that fold of 4 sequences
", "votes": null, "replies": [] }, { "id": 1691998, "author_name": "imeintanis", "author_url": "", "post_date": "02/15/2022 18:24:53", "content": "\n\nmy val set have 4 sequences: [26651, 59337, 53708, 45015]
\n
interesting!
\nour valid set consists of 5 seq., where 3/5 are similar to yours.
\nThis gave correlation with PVT ~90%
\nRef: 11th solution
@bakeryproducts cv score ranged around 0.78~0.8 @ 1.4x to 2.0x inference size. I the last month, I build a wide range solution to face the possible cots size variance in test set, but it is not the case finally.
\n@imeintanis interesting and congratulations π
", "votes": null, "replies": [] } ] }, { "id": 1690697, "author_name": "drzhuzhe", "author_url": "", "post_date": "02/15/2022 04:11:52", "content": "do you use Dynamichead or any other recently published detection head?
", "votes": null, "replies": [ { "id": 1690725, "author_name": "outrunner", "author_url": "", "post_date": "02/15/2022 04:30:24", "content": "No, just features -> conv -> conv -> heatmap
", "votes": null, "replies": [] }, { "id": 1691058, "author_name": "drzhuzhe", "author_url": "", "post_date": "02/15/2022 07:58:03", "content": "But, why 2 stage segment task can make such a improvement to detection?
\nI haven't seen this as a trick or methodβ¦.
base = keras_efficientnet_v2.EfficientNetV2B0(input_shape=image_size, drop_connect_rate=0, pretrained=\"imagenet21k\")
\nx = base.get_layer(\"stack_5_block0_sortcut_swish\").output
\nx = DilatedSpatialPyramidPooling(x)
input_a = UpSampling2D(size=(2, 2), interpolation=\"bilinear\")(x)
\ninput_b = base.get_layer('stack_3_block0_sortcut_swish').output
\ninput_b = convolution_block(input_b, num_filters=48, kernel_size=1)
features = Concatenate(axis=-1)([input_a, input_b])
\nx = convolution_block(features)
\nx = convolution_block(x)
\nheatmap = Conv2D(num_classes, 1, padding=\"same\", activation='sigmoid', name='hm', dtype=tf.float32)(x)
x = Conv2D(256, 3, padding=\"same\", activation='swish')(features)
\nx = Conv2D(256, 3, padding=\"same\", activation='swish')(x)
\nreg = Conv2D(4, 1, padding=\"same\", activation='linear', name='re', dtype=tf.float32)(x)
model = keras.Model(inputs=base.input, outputs={'hm':heatmap, 're':reg})
", "votes": null, "replies": [] }, { "id": 1691127, "author_name": "drzhuzhe", "author_url": "", "post_date": "02/15/2022 08:41:13", "content": "Thanks
\nbut what I most wonder is
\nFPN is also has a upsampling and feture concat,
\nI want to know why this methold have such a advantange than FPN ?
To my understand heatmap mask is generate by centernet
\nyou 2 stage model use centernet output and gt bbox as training target for make better use of object segmentation detail information
\nπ in next competition I will use this in my pipline too.
", "votes": null, "replies": [] }, { "id": 1691130, "author_name": "outrunner", "author_url": "", "post_date": "02/15/2022 08:45:30", "content": "I did not try other method, so I can not confirm whether it has advantage.
", "votes": null, "replies": [] } ] }, { "id": 1690711, "author_name": "lukaszborecki", "author_url": "", "post_date": "02/15/2022 04:20:47", "content": "Congratulations ! :) great work
", "votes": null, "replies": [ { "id": 1690726, "author_name": "outrunner", "author_url": "", "post_date": "02/15/2022 04:31:01", "content": "Thank you π
", "votes": null, "replies": [] } ] }, { "id": 1690721, "author_name": "sgalib", "author_url": "", "post_date": "02/15/2022 04:29:08", "content": "Congratulations. Winning gold medal requires trying new things and make it work. Thanks for sharing your solution!
", "votes": null, "replies": [ { "id": 1690730, "author_name": "outrunner", "author_url": "", "post_date": "02/15/2022 04:34:23", "content": "Thanks, I like objects as points method because it is easy to apply post processing.
", "votes": null, "replies": [] } ] }, { "id": 1690749, "author_name": "ivanlaulintiong", "author_url": "", "post_date": "02/15/2022 04:50:05", "content": "Congratulations. thanks for your sharing solution.
\nJust curious, is your solution implemented with TF Object Detection API? or in Keras + Tensorflow?
tf.keras, thanks.
", "votes": null, "replies": [] }, { "id": 1690863, "author_name": "dragonzhang", "author_url": "", "post_date": "02/15/2022 06:14:58", "content": "double winner!!!
", "votes": null, "replies": [] }, { "id": 1690895, "author_name": "outrunner", "author_url": "", "post_date": "02/15/2022 06:34:27", "content": "although I have model that can get 0.7+ within 45min, but I did not choose it. π
", "votes": null, "replies": [] } ] }, { "id": 1690901, "author_name": "remekkinas", "author_url": "", "post_date": "02/15/2022 06:40:15", "content": "Congratulations @outrunner Your scores (pub/priv) are outstanding. Everyday we checked LB looking for your improvements and β¦ asking what you did to cross 0.8 β¦ You are TOP Kaggler! Reading your solution description and I am impressed with the path you chose. ππππͺπͺπππ
", "votes": null, "replies": [ { "id": 1690936, "author_name": "outrunner", "author_url": "", "post_date": "02/15/2022 06:58:23", "content": "I have learned a lot from your sharing, thank you.
", "votes": null, "replies": [] } ] }, { "id": 1690941, "author_name": "lixxxxx", "author_url": "", "post_date": "02/15/2022 06:59:20", "content": "Good job bro!!!
", "votes": null, "replies": [] }, { "id": 1690962, "author_name": "cpmpml", "author_url": "", "post_date": "02/15/2022 07:09:08", "content": "Congrats on yet anogther solo gold!
", "votes": null, "replies": [ { "id": 1690989, "author_name": "outrunner", "author_url": "", "post_date": "02/15/2022 07:19:31", "content": "Thank you π
", "votes": null, "replies": [] } ] }, { "id": 1690976, "author_name": "bakeryproducts", "author_url": "", "post_date": "02/15/2022 07:14:01", "content": "@outrunner Congratulations! Turns out we were doing very similar thing, mine was centernet with unet, 1/4 input, decoupled head and many more tricks. I also tried EffNetV2 as backbone but NF-backbone was better on cv. Did you use separate offset head on 1/8? Did you train on full frames? How long did it take? I did it on crops, even 256x256 with specific sampling was enough.
\nNote to organizers: this setup can be easily ported to TRT to deploy on edge devices. I did it many times with lighter backbones on pretty much every jetson available (excluding Nano, including AGX). At the same time converting usual anchor-based net will be a nightmare.
", "votes": null, "replies": [ { "id": 1691023, "author_name": "outrunner", "author_url": "", "post_date": "02/15/2022 07:34:08", "content": "I train on some crop sizes to full frame randomly. For example, [high, width, batch_size] in [[448, 640, 20], [640, 896, 10], [720, 1280, 6]].
\nTraining takes about 1 day on single 2080ti.
Wow thats a long time. Same again, but I never went above 640, 896. What about offset head, i updated question a little
", "votes": null, "replies": [] }, { "id": 1691046, "author_name": "outrunner", "author_url": "", "post_date": "02/15/2022 07:48:53", "content": "x = Conv2D(256, 3)(features)
\nx = Conv2D(256, 3)(x)
\nreg = Conv2D(4, 1)(x)
\noffset + wh together π
Same, sorry for never-ending questions, As out pipelines are similar, im trying to understand was it bad coding or something else on my part. Did you do just regression on wh, like dirac-delta distribution, or something cooler like gauss, etc?
", "votes": null, "replies": [] }, { "id": 1691057, "author_name": "outrunner", "author_url": "", "post_date": "02/15/2022 07:57:29", "content": "oh, I see your update. offset head also is 1/8.
\nI post the model here
Thanks for the solution and congrats!
", "votes": null, "replies": [ { "id": 1692339, "author_name": "outrunner", "author_url": "", "post_date": "02/16/2022 01:30:29", "content": "Thank you π
", "votes": null, "replies": [] } ] }, { "id": 1691287, "author_name": "mednoun", "author_url": "", "post_date": "02/15/2022 10:33:25", "content": "Good job bro keep it up !!
", "votes": null, "replies": [] }, { "id": 1691316, "author_name": "kuanweichen", "author_url": "", "post_date": "02/15/2022 10:48:48", "content": "Congratulations! May I ask why you went with CenterNet instead of YOLO or fasterRCNN? Did you try multiple models out and found CenterNet gives the best result?
", "votes": null, "replies": [ { "id": 1692343, "author_name": "outrunner", "author_url": "", "post_date": "02/16/2022 01:34:34", "content": "No, I think anchor free base will perform well since the target is simple. (shape, texture, size, etc.)
", "votes": null, "replies": [] } ] }, { "id": 1691935, "author_name": "arunasivapragasam", "author_url": "", "post_date": "02/15/2022 17:36:04", "content": "congratulations @outrunner ππ
", "votes": null, "replies": [] }, { "id": 1692356, "author_name": "elldeejay", "author_url": "", "post_date": "02/16/2022 01:57:47", "content": "Great Works Keep it up!
", "votes": null, "replies": [] }, { "id": 1928051, "author_name": "onkur7", "author_url": "", "post_date": "09/06/2022 07:50:29", "content": "Is the code publicly available?
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1690578": "Since overfitting strategy failed, let's just talk about what I've done one month ago. π \n\nI use CenterNet with DeepLabV3+ architecture and EfficientNetV2 backbone, start from this [example](https://keras.io/examples/vision/deeplabv3_plus/).\n- change backbone to EfficientNetV2 B0~XL\n- change output heatmap size to 1/8 input, and add regression head\n- training on 1280x720, inference on 1792x1008(1.4x)\n- blend two can get private LB score 0.712, and 0.722 if inference on 1.6x\n\nActually, I am pleased to survive in huge shake up. π\n\nThanks everyone and congratulation to all winners.", "1690593": "Congratulations @outrunner", "1690594": "Simple description but definitely contains many experiences and thoughts inside! Thank you for sharing and amazing 1st place public LB position!", "1690597": "thank you π", "1690598": "I don't like data inconsistency, but you know, it is kaggle style.π ", "1690623": "Congrat for your money medal! Can you share your training code?", "1690634": "Thanks. Nothing special, just heavily augmentation, perspective transform, rotation, flip, mixup, random crop and batch size every step. Focal loss for heatmap, Huber loss for regression, and Adam optimizer.", "1690639": "a bit confused, `CenterNet` is for **Object Detection** and `DeepLabV3+` is for **Segmentation**. So, how did you use both of them?", "1690642": "You survived this shake up/down! I overfitted to public LB, but I got a bronze medal. Congrats on 4th place! Thanks for sharing your solution.", "1690643": "I don't know whether I should call it centernet, but objects as points model just like segmentation + regression head.", "1690644": "WoW. How did you get the mask?", "1690647": "Thanks, it's all luck. π ", "1690650": "reference CenterNet paper to gen the mask", "1690657": "I think there is also CenterNetV2. Did you try that also?", "1690663": "No, I did not try it.", "1690664": "Congratulation on Solo Win\nCan you talk about your validation strategy?", "1690666": "Thank youπ", "1690673": "as usual I use 1 fold.\nmy val set have 4 sequences: [26651, 59337, 53708, 45015]\nuse all cots frames and half none-cots frames for saving time.\nselect checkpoints by local F2 score and public LB score.", "1690690": "Thanks @outrunner How did you come about selecting those specific sequences", "1690697": "do you use Dynamichead or any other recently published detection head?", "1690701": "random selection according to mean and total cots counts, then decide by my mood. π", "1690711": "Congratulations ! :) great work", "1690721": "Congratulations. Winning gold medal requires trying new things and make it work. Thanks for sharing your solution!", "1690725": "No, just features -> conv -> conv -> heatmap", "1690726": "Thank you π", "1690730": "Thanks, I like objects as points method because it is easy to apply post processing.", "1690749": "Congratulations. thanks for your sharing solution.\nJust curious, is your solution implemented with TF Object Detection API? or in Keras + Tensorflow?", "1690752": "tf.keras, thanks.", "1690863": "double winner!!!", "1690895": "although I have model that can get 0.7+ within 45min, but I did not choose it. π ", "1690901": "Congratulations @outrunner Your scores (pub/priv) are outstanding. Everyday we checked LB looking for your improvements and ... asking what you did to cross 0.8 ... You are TOP Kaggler! Reading your solution description and I am impressed with the path you chose. ππππͺπͺπππ", "1690912": "outrunner I gotta learn this mood based decision from you π \nCongratulations !", "1690931": "haha, thanks. π", "1690936": "I have learned a lot from your sharing, thank you.", "1690941": "Good job bro!!!", "1690962": "Congrats on yet anogther solo gold!", "1690976": "outrunner Congratulations! Turns out we were doing very similar thing, mine was centernet with unet, 1/4 input, decoupled head and many more tricks. I also tried EffNetV2 as backbone but NF-backbone was better on cv. Did you use *separate* offset head on 1/8? Did you train on full frames? How long did it take? I did it on crops, even 256x256 with specific sampling was enough. \n\nNote to organizers: this setup can be easily ported to TRT to deploy on edge devices. I did it many times with lighter backbones on pretty much every jetson available (excluding Nano, including AGX). At the same time converting usual anchor-based net will be a nightmare.", "1690989": "Thank you π", "1690990": "awsaf49 I build second stage on top on centernet arch, not fully probabilistic as in CNV2 paper thou. It helps a little.", "1691023": "I train on some crop sizes to full frame randomly. For example, [high, width, batch_size] in [[448, 640, 20], [640, 896, 10], [720, 1280, 6]].\nTraining takes about 1 day on single 2080ti.", "1691035": "Wow thats a long time. Same again, but I never went above 640, 896. What about offset head, i updated question a little", "1691046": "` x = Conv2D(256, 3)(features)`\n` x = Conv2D(256, 3)(x)`\n` reg = Conv2D(4, 1)(x)`\noffset + wh together π", "1691052": "Same, sorry for never-ending questions, As out pipelines are similar, im trying to understand was it bad coding or something else on my part. Did you do just regression on wh, like dirac-delta distribution, or something cooler like gauss, etc?", "1691057": "oh, I see your update. offset head also is 1/8.\nI post the model [here](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/307626#1691065)", "1691058": "But, why 2 stage segment task can make such a improvement to detection?\nI haven't seen this as a trick or method....", "1691065": "base = keras_efficientnet_v2.EfficientNetV2B0(input_shape=image_size, drop_connect_rate=0, pretrained=\"imagenet21k\")\n\nx = base.get_layer(\"stack_5_block0_sortcut_swish\").output\nx = DilatedSpatialPyramidPooling(x)\n\ninput_a = UpSampling2D(size=(2, 2), interpolation=\"bilinear\")(x)\ninput_b = base.get_layer('stack_3_block0_sortcut_swish').output\ninput_b = convolution_block(input_b, num_filters=48, kernel_size=1)\n\nfeatures = Concatenate(axis=-1)([input_a, input_b])\n\nx = convolution_block(features)\nx = convolution_block(x)\nheatmap = Conv2D(num_classes, 1, padding=\"same\", activation='sigmoid', name='hm', dtype=tf.float32)(x)\n\nx = Conv2D(256, 3, padding=\"same\", activation='swish')(features)\nx = Conv2D(256, 3, padding=\"same\", activation='swish')(x)\nreg = Conv2D(4, 1, padding=\"same\", activation='linear', name='re', dtype=tf.float32)(x)\n\nmodel = keras.Model(inputs=base.input, outputs={'hm':heatmap, 're':reg})", "1691127": "Thanks \nbut what I most wonder is \nFPN is also has a upsampling and feture concat,\nI want to know why this methold have such a advantange than FPN ?\n\nTo my understand heatmap mask is generate by centernet\n\nyou 2 stage model use centernet output and gt bbox as training target for make better use of object segmentation detail information\n\nπ in next competition I will use this in my pipline too.", "1691130": "I did not try other method, so I can not confirm whether it has advantage.", "1691280": "Thanks for the solution and congrats!", "1691287": "Good job bro keep it up !!", "1691316": "Congratulations! May I ask why you went with CenterNet instead of YOLO or fasterRCNN? Did you try multiple models out and found CenterNet gives the best result?", "1691530": "Can you please share cv score on that fold of 4 sequences", "1691935": "congratulations @outrunner ππ", "1691998": ">my val set have 4 sequences: [26651, 59337, 53708, 45015]\n\ninteresting! \nour valid set consists of 5 seq., where 3/5 are similar to yours.\nThis gave correlation with PVT ~90% \nRef: [11th solution](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/307718)", "1692337": "bakeryproducts cv score ranged around 0.78~0.8 @ 1.4x to 2.0x inference size. I the last month, I build a wide range solution to face the possible cots size variance in test set, but it is not the case finally.\n\n@imeintanis interesting and congratulations π", "1692339": "Thank you π", "1692343": "No, I think anchor free base will perform well since the target is simple. (shape, texture, size, etc.)", "1692356": "Great Works Keep it up!", "1928051": "Is the code publicly available?" }, "source": "meta" }