![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F548996%2Fbb42658ab15675df44a4ab36aff46630%2Fmaterial_head.png?generation=1570245068374646&alt=media\")
Congrats to all the winners, and thanks to competition organizers for this interesting competition again.
\n\nI used almost same architecture as last year. \nSo I'll summarize only difference.
\n\nI made cascade-rcnn model using mmdetection.
\n\nmAP for 57 classes improved a lot, more than 0.1 compared to last year yolo model.
\n\nI made 3 models for this part, and then I made ensemble of them.
\n\nThis is the model I used for relation 'is' last year.
\n\nI made cascade-rcnn model which detect 42 'is-relation' classes.\nThis model is almost same as 'toshif' explained last year.
\n\nI added 'material' detection head to cascade-rcnn.\nThis model predict Bounding Box and class and material at the same time.
\n\n
Results:
\n\n|model |public |private |\n|---|---|---|\n|2-1-1 |0.07523 |0.07264 |\n|2-1-2 |0.08332 |0.08075 |\n|2-1-3 |0.08191 |0.07948 |\n|ensemble |0.08514 |0.08232 |
\n\nI expected 2-1-3 to have better score...
\n\nBase model is almost same as I shared here for this part.
\n\nI made expert models which only in charge of small sample class, and made ensemble of them with weighted average of their probability.
\n\nThis is the result AP for validation data:
\n\n|class |grand truth BB |predicted BB |\n|---|---|---|\n|at |93% |31% |\n|on |92% |32% |\n|holds |89% |54% |\n|plays |94% |58% |\n|interacts with |82% |45% |\n|inside of |72% |37% |\n|wears |94% |55% |\n|hits |55% |57% |\n|under |50% |20% |\n|mAP without hits/under |88% |45% |\n|mAP |80% |43% |
\n\nFor grand truth BB pairs, this relationships prediction model has very high accuracy.\nmAP without hits/under which have very small samples is 88%!
\n\nI did not used Light GBM for this part.\nI just used simple formula.
\n\nFinal Score = Object1Score x Object2Score x RelationsipScore
This year, my LB score improved to 0.38818 from last year score 0.23709.\nMost of this improvement comes from object detection improvement.
\n\nIt seems that good object detection is the most important part of this competition.
\n\nBTW, I became GM as of this competition. \nI'd like to thank to my previous team mate. \nI learned lots of things from them and I could not be GM without them.\nThank you, Carl, Little Boat, KazAnova, Ahmet, Kohei-san, Akiyama-san, owruby!
", "messages": [ { "id": "641744", "postDate": "10/05/2019 03:27:37", "content": "Congrats to all the winners, and thanks to competition organizers for this interesting competition again.
\n\nI used almost same architecture as last year. \nSo I'll summarize only difference.
\n\nI made cascade-rcnn model using mmdetection.
\n\nmAP for 57 classes improved a lot, more than 0.1 compared to last year yolo model.
\n\nI made 3 models for this part, and then I made ensemble of them.
\n\nThis is the model I used for relation 'is' last year.
\n\nI made cascade-rcnn model which detect 42 'is-relation' classes.\nThis model is almost same as 'toshif' explained last year.
\n\nI added 'material' detection head to cascade-rcnn.\nThis model predict Bounding Box and class and material at the same time.
\n\n
Results:
\n\n|model |public |private |\n|---|---|---|\n|2-1-1 |0.07523 |0.07264 |\n|2-1-2 |0.08332 |0.08075 |\n|2-1-3 |0.08191 |0.07948 |\n|ensemble |0.08514 |0.08232 |
\n\nI expected 2-1-3 to have better score...
\n\nBase model is almost same as I shared here for this part.
\n\nI made expert models which only in charge of small sample class, and made ensemble of them with weighted average of their probability.
\n\nThis is the result AP for validation data:
\n\n|class |grand truth BB |predicted BB |\n|---|---|---|\n|at |93% |31% |\n|on |92% |32% |\n|holds |89% |54% |\n|plays |94% |58% |\n|interacts with |82% |45% |\n|inside of |72% |37% |\n|wears |94% |55% |\n|hits |55% |57% |\n|under |50% |20% |\n|mAP without hits/under |88% |45% |\n|mAP |80% |43% |
\n\nFor grand truth BB pairs, this relationships prediction model has very high accuracy.\nmAP without hits/under which have very small samples is 88%!
\n\nI did not used Light GBM for this part.\nI just used simple formula.
\n\nFinal Score = Object1Score x Object2Score x RelationsipScore
This year, my LB score improved to 0.38818 from last year score 0.23709.\nMost of this improvement comes from object detection improvement.
\n\nIt seems that good object detection is the most important part of this competition.
\n\nBTW, I became GM as of this competition. \nI'd like to thank to my previous team mate. \nI learned lots of things from them and I could not be GM without them.\nThank you, Carl, Little Boat, KazAnova, Ahmet, Kohei-san, Akiyama-san, owruby!
", "rawMarkdown": "Congrats to all the winners, and thanks to competition organizers for this interesting competition again.\n\nI used almost same architecture as [last year](https://www.kaggle.com/c/google-ai-open-images-visual-relationship-track/discussion/64651). \nSo I'll summarize only difference.\n\n\n# Model 1: object detection\nI made cascade-rcnn model using [mmdetection](https://github.com/open-mmlab/mmdetection).\n\nmAP for 57 classes improved a lot, more than 0.1 compared to last year yolo model.\n\n\n\n\n# Model 2: visual relationship\n\n## 2-1: relation 'is'\nI made 3 models for this part, and then I made ensemble of them.\n\n\n### 2-1-1: relation 'is' (2 stage model)\nThis is the model I used for relation 'is' last year.\n\n\n### 2-1-2: relation 'is' (1 stage model)\nI made cascade-rcnn model which detect 42 'is-relation' classes.\nThis model is almost same as ['toshif' explained last year](https://www.kaggle.com/c/google-ai-open-images-visual-relationship-track/discussion/64642).\n\n### 2-1-3: relation 'is' (1 stage model with material head)\nI added 'material' detection head to cascade-rcnn.\nThis model predict Bounding Box and class and material at the same time.\n\n\n\nResults:\n\n|model |public |private |\n|---|---|---|\n|2-1-1 |0.07523 |0.07264 |\n|2-1-2 |0.08332 |0.08075 |\n|2-1-3 |0.08191 |0.07948 |\n|ensemble |0.08514 |0.08232 |\n\nI expected 2-1-3 to have better score...\n\n\n\n## 2-2: Triplet Relationships\nBase model is almost same as I shared [here](https://www.kaggle.com/c/google-ai-open-images-visual-relationship-track/discussion/64651#380288) for this part.\n\n\nI made expert models which only in charge of small sample class, and made ensemble of them with weighted average of their probability.\n\nThis is the result AP for validation data:\n\n|class |grand truth BB |predicted BB |\n|---|---|---|\n|at |93% |31% |\n|on |92% |32% |\n|holds |89% |54% |\n|plays |94% |58% |\n|interacts with |82% |45% |\n|inside of |72% |37% |\n|wears |94% |55% |\n|hits |55% |57% |\n|under |50% |20% |\n|mAP without hits/under |88% |45% |\n|mAP |80% |43% |\n\nFor grand truth BB pairs, this relationships prediction model has very high accuracy.\nmAP without hits/under which have very small samples is 88%!\n\n# Model 3: Final Score Prediction\nI did not used Light GBM for this part.\nI just used simple formula.\n\n`Final Score = Object1Score x Object2Score x RelationsipScore`\n\n\n\nThis year, my LB score improved to 0.38818 from last year score 0.23709.\nMost of this improvement comes from object detection improvement.\n\nIt seems that good object detection is the most important part of this competition.\n\n\n\n\n\nBTW, I became GM as of this competition. \nI'd like to thank to my previous team mate. \nI learned lots of things from them and I could not be GM without them.\nThank you, Carl, Little Boat, KazAnova, Ahmet, Kohei-san, Akiyama-san, owruby!", "votes": null }, { "id": "641776", "postDate": "10/05/2019 04:49:13", "content": "Congratulations\nGreat Write-Up\nThank You for Sharing Your Approach & Insights….!! @its7171
", "rawMarkdown": "Congratulations\nGreat Write-Up\nThank You for Sharing Your Approach & Insights….!! @its7171", "votes": null }, { "id": "641958", "postDate": "10/05/2019 11:25:21", "content": "Congratulations. How long did it take to train cascade-rcnn?
", "rawMarkdown": "Congratulations. How long did it take to train cascade-rcnn?", "votes": null }, { "id": "641973", "postDate": "10/05/2019 11:50:29", "content": "thanks, it took about 2 weeks with 2 x 1080ti.
", "rawMarkdown": "thanks, it took about 2 weeks with 2 x 1080ti.", "votes": null }, { "id": "643823", "postDate": "10/08/2019 01:00:00", "content": "Congratulations on new GM!!\nThank you for sharing your approach. It is so great.
", "rawMarkdown": "Congratulations on new GM!!\nThank you for sharing your approach. It is so great.", "votes": null }, { "id": "644503", "postDate": "10/08/2019 22:44:27", "content": "Nice presentation. great achievement as GM\nkeep it up.\nA
", "rawMarkdown": "Nice presentation. great achievement as GM\nkeep it up.\nA", "votes": null } ], "comments": [ { "id": 641776, "author_name": "veeralakrishna", "author_url": "", "post_date": "10/05/2019 04:49:13", "content": "Congratulations\nGreat Write-Up\nThank You for Sharing Your Approach & Insights….!! @its7171
", "votes": null, "replies": [] }, { "id": 641958, "author_name": "zgkkgyy", "author_url": "", "post_date": "10/05/2019 11:25:21", "content": "Congratulations. How long did it take to train cascade-rcnn?
", "votes": null, "replies": [ { "id": 641973, "author_name": "its7171", "author_url": "", "post_date": "10/05/2019 11:50:29", "content": "thanks, it took about 2 weeks with 2 x 1080ti.
", "votes": null, "replies": [] } ] }, { "id": 643823, "author_name": "owruby", "author_url": "", "post_date": "10/08/2019 01:00:00", "content": "Congratulations on new GM!!\nThank you for sharing your approach. It is so great.
", "votes": null, "replies": [] }, { "id": 644503, "author_name": "zinovadr", "author_url": "", "post_date": "10/08/2019 22:44:27", "content": "Nice presentation. great achievement as GM\nkeep it up.\nA
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "641744": "Congrats to all the winners, and thanks to competition organizers for this interesting competition again.\n\nI used almost same architecture as [last year](https://www.kaggle.com/c/google-ai-open-images-visual-relationship-track/discussion/64651). \nSo I'll summarize only difference.\n\n\n# Model 1: object detection\nI made cascade-rcnn model using [mmdetection](https://github.com/open-mmlab/mmdetection).\n\nmAP for 57 classes improved a lot, more than 0.1 compared to last year yolo model.\n\n\n\n\n# Model 2: visual relationship\n\n## 2-1: relation 'is'\nI made 3 models for this part, and then I made ensemble of them.\n\n\n### 2-1-1: relation 'is' (2 stage model)\nThis is the model I used for relation 'is' last year.\n\n\n### 2-1-2: relation 'is' (1 stage model)\nI made cascade-rcnn model which detect 42 'is-relation' classes.\nThis model is almost same as ['toshif' explained last year](https://www.kaggle.com/c/google-ai-open-images-visual-relationship-track/discussion/64642).\n\n### 2-1-3: relation 'is' (1 stage model with material head)\nI added 'material' detection head to cascade-rcnn.\nThis model predict Bounding Box and class and material at the same time.\n\n\n\nResults:\n\n|model |public |private |\n|---|---|---|\n|2-1-1 |0.07523 |0.07264 |\n|2-1-2 |0.08332 |0.08075 |\n|2-1-3 |0.08191 |0.07948 |\n|ensemble |0.08514 |0.08232 |\n\nI expected 2-1-3 to have better score...\n\n\n\n## 2-2: Triplet Relationships\nBase model is almost same as I shared [here](https://www.kaggle.com/c/google-ai-open-images-visual-relationship-track/discussion/64651#380288) for this part.\n\n\nI made expert models which only in charge of small sample class, and made ensemble of them with weighted average of their probability.\n\nThis is the result AP for validation data:\n\n|class |grand truth BB |predicted BB |\n|---|---|---|\n|at |93% |31% |\n|on |92% |32% |\n|holds |89% |54% |\n|plays |94% |58% |\n|interacts with |82% |45% |\n|inside of |72% |37% |\n|wears |94% |55% |\n|hits |55% |57% |\n|under |50% |20% |\n|mAP without hits/under |88% |45% |\n|mAP |80% |43% |\n\nFor grand truth BB pairs, this relationships prediction model has very high accuracy.\nmAP without hits/under which have very small samples is 88%!\n\n# Model 3: Final Score Prediction\nI did not used Light GBM for this part.\nI just used simple formula.\n\n`Final Score = Object1Score x Object2Score x RelationsipScore`\n\n\n\nThis year, my LB score improved to 0.38818 from last year score 0.23709.\nMost of this improvement comes from object detection improvement.\n\nIt seems that good object detection is the most important part of this competition.\n\n\n\n\n\nBTW, I became GM as of this competition. \nI'd like to thank to my previous team mate. \nI learned lots of things from them and I could not be GM without them.\nThank you, Carl, Little Boat, KazAnova, Ahmet, Kohei-san, Akiyama-san, owruby!", "641776": "Congratulations\nGreat Write-Up\nThank You for Sharing Your Approach & Insights….!! @its7171", "641958": "Congratulations. How long did it take to train cascade-rcnn?", "641973": "thanks, it took about 2 weeks with 2 x 1080ti.", "643823": "Congratulations on new GM!!\nThank you for sharing your approach. It is so great.", "644503": "Nice presentation. great achievement as GM\nkeep it up.\nA" }, "source": "meta" }