![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1146523%2Fe323a35a9255e8e50131807aa19b56ef%2Fsolution.png?generation=1677715798541475&alt=media\")
First of all, I would like to thank our hosts for organizing the competition.
\nIt was a task I've never solved before, and it was both educational and a lot of fun trying different approaches!
\nx_contact = model_contact(x) \nx_g = model_g(x) \nnot_is_g = (is_g == )\nx = x_contact * not_is_g + x_g * is_g \n\n ():\n x_sideline = cnn(x_sideline_image) \n x_endzone = cnn(x_endzone_image)\n fc(torch.cat([x_sideline, x_endzone])), fc_sideline(x_sideline), fc_endzone(x_endzone)\nWeighted ensemble, G and contact respectively.
\n bbox_left_ratio = \n bbox_right_ratio = \n bbox_top_ratio = \n bbox_down_ratio = \n col [, , , ]:\n df[col] = df[[, ]].mean(axis=)\n df[] = df[[, ]].mean(axis=)\n df[] = df.groupby([, , ])[].transform()\n\n series = df.iloc[] \n left = (series[] - series[] * bbox_left_ratio)\n right = (series[] + series[] * bbox_right_ratio)\n top = (series[] + series[] * bbox_top_ratio)\n down = (series[] - series[] * bbox_down_ratio)\n img = img[down:top, left:right]\n img = cv2.resize(img, (, ))\nFirst of all, I would like to thank our hosts for organizing the competition.
\nIt was a task I've never solved before, and it was both educational and a lot of fun trying different approaches!
\nx_contact = model_contact(x) \nx_g = model_g(x) \nnot_is_g = (is_g == )\nx = x_contact * not_is_g + x_g * is_g \n\n ():\n x_sideline = cnn(x_sideline_image) \n x_endzone = cnn(x_endzone_image)\n fc(torch.cat([x_sideline, x_endzone])), fc_sideline(x_sideline), fc_endzone(x_endzone)\nWeighted ensemble, G and contact respectively.
\n bbox_left_ratio = \n bbox_right_ratio = \n bbox_top_ratio = \n bbox_down_ratio = \n col [, , , ]:\n df[col] = df[[, ]].mean(axis=)\n df[] = df[[, ]].mean(axis=)\n df[] = df.groupby([, , ])[].transform()\n\n series = df.iloc[] \n left = (series[] - series[] * bbox_left_ratio)\n right = (series[] + series[] * bbox_right_ratio)\n top = (series[] + series[] * bbox_top_ratio)\n down = (series[] - series[] * bbox_down_ratio)\n img = img[down:top, left:right]\n img = cv2.resize(img, (, ))\nFirst of all congratulations, thanks for sharing this great solution. I have some questions that in 3D CNN you predict 19 steps, that is, given that input sequence you will predict which step it will belong to? (classification)
", "rawMarkdown": "First of all congratulations, thanks for sharing this great solution. I have some questions that in 3D CNN you predict 19 steps, that is, given that input sequence you will predict which step it will belong to? (classification)", "votes": null }, { "id": "2168963", "postDate": "03/04/2023 16:59:24", "content": "Great work @kurupical and thanks for sharing this solution write up.
", "rawMarkdown": "Great work @kurupical and thanks for sharing this solution write up.", "votes": null } ], "comments": [ { "id": 2165136, "author_name": "nhtnguyntrng", "author_url": "", "post_date": "03/02/2023 02:29:05", "content": "First of all congratulations, thanks for sharing this great solution. I have some questions that in 3D CNN you predict 19 steps, that is, given that input sequence you will predict which step it will belong to? (classification)
", "votes": null, "replies": [] }, { "id": 2168963, "author_name": "robikscube", "author_url": "", "post_date": "03/04/2023 16:59:24", "content": "Great work @kurupical and thanks for sharing this solution write up.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2165032": "First of all, I would like to thank our hosts for organizing the competition.\nIt was a task I've never solved before, and it was both educational and a lot of fun trying different approaches!\n\n# Summary\n\n\n# Model Detail\n## 3D-CNN (cv: 0.770)\n- backbone: r3d_18 (from torchvision: https://pytorch.org/vision/stable/models/generated/torchvision.models.video.r3d_18.html#torchvision.models.video.R3D_18_Weights)\n- use 63 frames(20fps)\n- predict 19 steps\n- train every 9 steps\n- StepLR Scheduler(~2epochs: lr=1e-3/1e-4)\n\n## 2.5D3D-CNN (cv: 0.768)\n- Almost same as DFL's 1st solution by Team Hydrogen (https://www.kaggle.com/competitions/dfl-bundesliga-data-shootout/discussion/359932)\n- backbone: legacy_seresnet34\n- use 123 frames(20fps)\n- predict 3 frames\n- down sampling (g: 10%, contact: 30%)\n- label smoothing (0.1-0.9)\n\n## Both 3D, 2.5D3D\n- Linear layers for g and contact\n```python\n# x: (bs, cnn_features)\nx_contact = model_contact(x) # (bs, n_predict_frames)\nx_g = model_g(x) # (bs, n_predict_frames)\nnot_is_g = (is_g == 0)\nx = x_contact * not_is_g + x_g * is_g # (bs, n_predict_frames)\n```\n- output 3 prediction and calculate loss: only sideline, only endzone, concat sideline-endzone feature.\n```python\n# pseudo code\ndef forward(self, x_sideline_image, x_endzone_image):\n x_sideline = cnn(x_sideline_image) \n x_endzone = cnn(x_endzone_image)\n return fc(torch.cat([x_sideline, x_endzone])), fc_sideline(x_sideline), fc_endzone(x_endzone)\n```\n\n## LGBM (cv: 0.740)\n- about 1100 features\n- feautres\n - player's distance (tracking, helmet)\n - lag, diff\n - top_n nearest player's distance (n: parameters)\n - number of people within distance n (n: parameters)\n- groupby\n - game_play\n - is_g\n - is_same_team\n - number of people within distance n \n\n## ensemble\nWeighted ensemble, G and contact respectively.\n\n## What worked for me\n- image preprocessing\n - draw bbox -> draw bbox and paint out\n - use 2 colors(g, contact) -> use 3 colors(g, same team contact, different team contact)\n - crop the image with keeping the aspect ratio\n ```python\n bbox_left_ratio = 4.5\n bbox_right_ratio = 4.5\n bbox_top_ratio = 4.5\n bbox_down_ratio = 2.25\n for col in [\"x\", \"y\", \"width\", \"height\"]:\n df[col] = df[[f\"{col}_1\", f\"{col}_2\"]].mean(axis=1)\n df[\"bbox_size\"] = df[[\"width\", \"height\"]].mean(axis=1)\n df[\"bbox_size\"] = df.groupby([\"view\", \"step\", \"game_play\"])[\"bbox_size\"].transform(\"mean\")\n\n series = df.iloc[0] # sample\n left = int(series[\"x\"] - series[\"bbox_size\"] * bbox_left_ratio)\n right = int(series[\"x\"] + series[\"bbox_size\"] * bbox_right_ratio)\n top = int(series[\"y\"] + series[\"bbox_size\"] * bbox_top_ratio)\n down = int(series[\"y\"] - series[\"bbox_size\"] * bbox_down_ratio)\n img = img[down:top, left:right]\n img = cv2.resize(img, (128, 96))\n ```\n \n- StepLR with warmup scheduler\n- label smoothing (worked for 2.5D3D, but not worked for 3D)\n\n## What not worked for me\n- Transformers\n - use top 100~400 features of lgbm feature importances\n - tuned hard but got cv 0.02 lower than lgbm.\n- 2D->1D CNN\n - contact score is same as 2.5D3D, 3D but very poor G score in my work.\n- interpolate bbox\n\n\n## Other\n- tools: I make tools to investigate wrong inference and make a hypothesize to improve score.\nhttps://github.com/kurupical/nfl_contact_detection/blob/master/58218_003210_contact_0.506591796875_score0.0_H23_V10.gif", "2165136": "First of all congratulations, thanks for sharing this great solution. I have some questions that in 3D CNN you predict 19 steps, that is, given that input sequence you will predict which step it will belong to? (classification)", "2168963": "Great work @kurupical and thanks for sharing this solution write up." }, "source": "meta" }