Hi,\nI am a beginner in computer vision and object detection. I have some experience in making image classifiers in Keras using transfer learning but object detection is completely new for me. This competition seemed interesting so I started working on it, exploring similar past competitions on Kaggle.
\n\nFrom what I have learned, I think the following steps need to be followed:\n1. Cropping the characters using annotations given, from the train images.\n2. training a CNN based on these cropped images.\n3. Dividing the test images into boxes and using every box as an image for our CNN model.
\n\nIs this the approach or am I missing something here ? \nAlso, which object detection models(Yolo, R-CNN ) are good and easy to implement ? \nIf you can share any simple implementation of object detection models on custom datasets like this, that'd be great !
", "messages": [ { "id": "609526", "postDate": "08/27/2019 20:04:45", "content": "Hi,\nI am a beginner in computer vision and object detection. I have some experience in making image classifiers in Keras using transfer learning but object detection is completely new for me. This competition seemed interesting so I started working on it, exploring similar past competitions on Kaggle.
\n\nFrom what I have learned, I think the following steps need to be followed:\n1. Cropping the characters using annotations given, from the train images.\n2. training a CNN based on these cropped images.\n3. Dividing the test images into boxes and using every box as an image for our CNN model.
\n\nIs this the approach or am I missing something here ? \nAlso, which object detection models(Yolo, R-CNN ) are good and easy to implement ? \nIf you can share any simple implementation of object detection models on custom datasets like this, that'd be great !
", "rawMarkdown": "Hi,\nI am a beginner in computer vision and object detection. I have some experience in making image classifiers in Keras using transfer learning but object detection is completely new for me. This competition seemed interesting so I started working on it, exploring similar past competitions on Kaggle. \n\nFrom what I have learned, I think the following steps need to be followed:\n1. Cropping the characters using annotations given, from the train images.\n2. training a CNN based on these cropped images.\n3. Dividing the test images into boxes and using every box as an image for our CNN model.\n\nIs this the approach or am I missing something here ? \nAlso, which object detection models(Yolo, R-CNN ) are good and easy to implement ? \nIf you can share any simple implementation of object detection models on custom datasets like this, that'd be great !", "votes": null }, { "id": "609834", "postDate": "08/28/2019 06:29:15", "content": "Hi,
\n\nThe approach you are taking is the old approach where a sliding window was used to determine the location of the object.
\n\nYou should use the newer approaches like YOLO and R-CNN where the CNN works like a Sliding window.\nThe network output contains both the location and classification of the object in a structure implemented in CNN Dimensions.
\n\nFor example, In YOLO's Last layer (depends on implementation) which is a (7x7x30) CNN layer, the 7X7 means 49 grid cells. Each grid cell has 30 values from CNN's 3rd Dimension. The 30 values have structured meaning, where values form a structure to predict object bounding box size, location and object classifications.
\n\nHope this Helps
", "rawMarkdown": "Hi,\n\nThe approach you are taking is the old approach where a **sliding window** was used to determine the location of the object.\n\nYou should use the newer approaches like YOLO and R-CNN where the CNN works like a Sliding window.\nThe network output contains both the location and classification of the object in a structure implemented in CNN Dimensions.\n \nFor example, In YOLO's Last layer *(depends on implementation)* which is a **(7x7x30)** CNN layer, the 7X7 means 49 grid cells. **Each grid cell has 30 values** from CNN's 3rd Dimension. The 30 values have *structured meaning,* where values form a structure to *predict object bounding box size, location and object classifications*.\n\nHope this Helps", "votes": null }, { "id": "609996", "postDate": "08/28/2019 09:59:40", "content": "But YOLO is trained on some predefined classes. How do we train it on custom dataset ?\nDo we have to create a txt file for every image in the dataset, and then write the class and location of the objects into it ?
", "rawMarkdown": "But YOLO is trained on some predefined classes. How do we train it on custom dataset ?\nDo we have to create a txt file for every image in the dataset, and then write the class and location of the objects into it ?", "votes": null }, { "id": "610237", "postDate": "08/28/2019 15:10:51", "content": "Yes, If you want to train on custom dataset, you have to give the model with the location of every training object and the image to the model. \nhere is an example dataset for detection: https://github.com/moskewcz/VOCdevkit/tree/master/VOC2007
\n\nYou can use DarkFlow implementation(https://github.com/thtrieu/darkflow) of YOLO to train with this format.\nYou can also use an Annotation tool to annotate and get the data in the right format.
", "rawMarkdown": "Yes, If you want to **train on custom dataset**, you have to give the model with the **location of every training object and the image** to the model. \nhere is an **example dataset** for detection: https://github.com/moskewcz/VOCdevkit/tree/master/VOC2007\n\nYou can use **DarkFlow implementation**(https://github.com/thtrieu/darkflow) of YOLO to train with this format.\nYou can also use an Annotation tool to annotate and get the data in the right format.", "votes": null }, { "id": "611478", "postDate": "08/29/2019 09:27:18", "content": "Thanks a lot for these resources.
", "rawMarkdown": "Thanks a lot for these resources.", "votes": null }, { "id": "613014", "postDate": "08/30/2019 06:55:44", "content": "me too.
", "rawMarkdown": "me too.", "votes": null }, { "id": "629908", "postDate": "09/19/2019 13:01:40", "content": "Excuse me, so the input to the model will be the whole image, and the labels and bboxes for every character?? the same form of train.csv?
", "rawMarkdown": "Excuse me, so the input to the model will be the whole image, and the labels and bboxes for every character?? the same form of train.csv?", "votes": null }, { "id": "630176", "postDate": "09/19/2019 20:44:48", "content": "Yes, one possible approach is to treat this as object detection, then the inputs and outputs are exactly as you described.
", "rawMarkdown": "Yes, one possible approach is to treat this as object detection, then the inputs and outputs are exactly as you described.", "votes": null }, { "id": "630754", "postDate": "09/20/2019 17:10:05", "content": "@hassanalsamahi The input to the model is a resized image (size depends on model) and output is labels and object box coordinates prediction(for models like YOLO) for a set max number of objects.
\n\nFor training object detection models you may have to give both label and object box coordinates to the model. so that model can learn from the data.
", "rawMarkdown": "hassanalsamahi The input to the model is a **resized image** (size depends on model) and output is **labels and object box coordinates prediction**(for models like YOLO) for a set max number of objects.\n\nFor **training** object detection models you may have to give **both label and object box coordinates** to the model. so that model can learn from the data.", "votes": null }, { "id": "630780", "postDate": "09/20/2019 17:57:39", "content": "Okay great, thanks for your explanation
", "rawMarkdown": "Okay great, thanks for your explanation", "votes": null }, { "id": "630788", "postDate": "09/20/2019 18:06:36", "content": "@lopuhin thanks for your help
", "rawMarkdown": "lopuhin thanks for your help", "votes": null } ], "comments": [ { "id": 609834, "author_name": "rameshkamath", "author_url": "", "post_date": "08/28/2019 06:29:15", "content": "Hi,
\n\nThe approach you are taking is the old approach where a sliding window was used to determine the location of the object.
\n\nYou should use the newer approaches like YOLO and R-CNN where the CNN works like a Sliding window.\nThe network output contains both the location and classification of the object in a structure implemented in CNN Dimensions.
\n\nFor example, In YOLO's Last layer (depends on implementation) which is a (7x7x30) CNN layer, the 7X7 means 49 grid cells. Each grid cell has 30 values from CNN's 3rd Dimension. The 30 values have structured meaning, where values form a structure to predict object bounding box size, location and object classifications.
\n\nHope this Helps
", "votes": null, "replies": [ { "id": 609996, "author_name": "harshitt21", "author_url": "", "post_date": "08/28/2019 09:59:40", "content": "But YOLO is trained on some predefined classes. How do we train it on custom dataset ?\nDo we have to create a txt file for every image in the dataset, and then write the class and location of the objects into it ?
", "votes": null, "replies": [] }, { "id": 610237, "author_name": "rameshkamath", "author_url": "", "post_date": "08/28/2019 15:10:51", "content": "Yes, If you want to train on custom dataset, you have to give the model with the location of every training object and the image to the model. \nhere is an example dataset for detection: https://github.com/moskewcz/VOCdevkit/tree/master/VOC2007
\n\nYou can use DarkFlow implementation(https://github.com/thtrieu/darkflow) of YOLO to train with this format.\nYou can also use an Annotation tool to annotate and get the data in the right format.
", "votes": null, "replies": [] }, { "id": 611478, "author_name": "harshitt21", "author_url": "", "post_date": "08/29/2019 09:27:18", "content": "Thanks a lot for these resources.
", "votes": null, "replies": [] }, { "id": 629908, "author_name": "hassanalsamahi", "author_url": "", "post_date": "09/19/2019 13:01:40", "content": "Excuse me, so the input to the model will be the whole image, and the labels and bboxes for every character?? the same form of train.csv?
", "votes": null, "replies": [] }, { "id": 630176, "author_name": "lopuhin", "author_url": "", "post_date": "09/19/2019 20:44:48", "content": "Yes, one possible approach is to treat this as object detection, then the inputs and outputs are exactly as you described.
", "votes": null, "replies": [] }, { "id": 630754, "author_name": "rameshkamath", "author_url": "", "post_date": "09/20/2019 17:10:05", "content": "@hassanalsamahi The input to the model is a resized image (size depends on model) and output is labels and object box coordinates prediction(for models like YOLO) for a set max number of objects.
\n\nFor training object detection models you may have to give both label and object box coordinates to the model. so that model can learn from the data.
", "votes": null, "replies": [] }, { "id": 630780, "author_name": "hassanalsamahi", "author_url": "", "post_date": "09/20/2019 17:57:39", "content": "Okay great, thanks for your explanation
", "votes": null, "replies": [] }, { "id": 630788, "author_name": "hassanalsamahi", "author_url": "", "post_date": "09/20/2019 18:06:36", "content": "@lopuhin thanks for your help
", "votes": null, "replies": [] } ] }, { "id": 613014, "author_name": "need4data", "author_url": "", "post_date": "08/30/2019 06:55:44", "content": "me too.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "609526": "Hi,\nI am a beginner in computer vision and object detection. I have some experience in making image classifiers in Keras using transfer learning but object detection is completely new for me. This competition seemed interesting so I started working on it, exploring similar past competitions on Kaggle. \n\nFrom what I have learned, I think the following steps need to be followed:\n1. Cropping the characters using annotations given, from the train images.\n2. training a CNN based on these cropped images.\n3. Dividing the test images into boxes and using every box as an image for our CNN model.\n\nIs this the approach or am I missing something here ? \nAlso, which object detection models(Yolo, R-CNN ) are good and easy to implement ? \nIf you can share any simple implementation of object detection models on custom datasets like this, that'd be great !", "609834": "Hi,\n\nThe approach you are taking is the old approach where a **sliding window** was used to determine the location of the object.\n\nYou should use the newer approaches like YOLO and R-CNN where the CNN works like a Sliding window.\nThe network output contains both the location and classification of the object in a structure implemented in CNN Dimensions.\n \nFor example, In YOLO's Last layer *(depends on implementation)* which is a **(7x7x30)** CNN layer, the 7X7 means 49 grid cells. **Each grid cell has 30 values** from CNN's 3rd Dimension. The 30 values have *structured meaning,* where values form a structure to *predict object bounding box size, location and object classifications*.\n\nHope this Helps", "609996": "But YOLO is trained on some predefined classes. How do we train it on custom dataset ?\nDo we have to create a txt file for every image in the dataset, and then write the class and location of the objects into it ?", "610237": "Yes, If you want to **train on custom dataset**, you have to give the model with the **location of every training object and the image** to the model. \nhere is an **example dataset** for detection: https://github.com/moskewcz/VOCdevkit/tree/master/VOC2007\n\nYou can use **DarkFlow implementation**(https://github.com/thtrieu/darkflow) of YOLO to train with this format.\nYou can also use an Annotation tool to annotate and get the data in the right format.", "611478": "Thanks a lot for these resources.", "613014": "me too.", "629908": "Excuse me, so the input to the model will be the whole image, and the labels and bboxes for every character?? the same form of train.csv?", "630176": "Yes, one possible approach is to treat this as object detection, then the inputs and outputs are exactly as you described.", "630754": "hassanalsamahi The input to the model is a **resized image** (size depends on model) and output is **labels and object box coordinates prediction**(for models like YOLO) for a set max number of objects.\n\nFor **training** object detection models you may have to give **both label and object box coordinates** to the model. so that model can learn from the data.", "630780": "Okay great, thanks for your explanation", "630788": "lopuhin thanks for your help" }, "source": "meta" }