{ "id": 541458, "title": "FLAME AI Second Place Solution", "content": "

First of all, thank you to the hosts for organizing this competition. It was a truly great experience competing with all of you!

\n

My Solution
\nFeature Engineering
\nInclusion of Fire-Line Differences:
\nI computed the difference between the fire-line at the current timestep and the previous timestep

\n

Inclusion of Surrounding Positions:
\nTo improve spatial awareness, I included information from surrounding positions.

\n

Model Architecture
\nThe model architecture consists of four main parts: CNN -> ConvLSTM -> Transformer -> Decoder.

\n

CNN
\nThe CNN handles spatial feature extraction with two branches:

\n

Local Branch: This branch captures fine-grained, localized information from the fire-line.
\nGlobal Branch: Inspired by U-Net, this branch captures broader spatial context.
\nBoth branches are concatenated and passed through a convolution block for further processing.
\nI used a MBConvolution-like block with ECA (Efficient Channel Attention) and spatial attention.

\n

ConvLSTM
\nThe ConvLSTM layer plays a crucial role as the main autoregressive component, predicting the next frame in the form of a latent vector.

\n

Transformer
\nThe Transformer refines the latent vector by leveraging attention mechanisms, enabling the model to capture long-range temporal dependencies.

\n

Decoder
\nThe Decoder takes the refined latent vector and outputs the predicted fire-line for the next timestep.

\n

Edit:
\nTraining notebook: https://www.kaggle.com/code/jobayerhossain/flame-ai-training/notebook
\nInference notebook: https://www.kaggle.com/code/jobayerhossain/flameai-inference/notebook

", "comments": [ { "id": 3076723, "author_name": "Deepak Saldanha", "author_url": "", "post_date": "2024-12-20T08:05:53.807000", "content": "

Many Congratulations and thank you for sharing your work!

", "votes": 0, "replies": [] } ], "comment_count": 1, "messages": [ { "id": 3022535, "postDate": "2024-10-19T17:00:55.137Z", "content": "

First of all, thank you to the hosts for organizing this competition. It was a truly great experience competing with all of you!

\n

My Solution
\nFeature Engineering
\nInclusion of Fire-Line Differences:
\nI computed the difference between the fire-line at the current timestep and the previous timestep

\n

Inclusion of Surrounding Positions:
\nTo improve spatial awareness, I included information from surrounding positions.

\n

Model Architecture
\nThe model architecture consists of four main parts: CNN -> ConvLSTM -> Transformer -> Decoder.

\n

CNN
\nThe CNN handles spatial feature extraction with two branches:

\n

Local Branch: This branch captures fine-grained, localized information from the fire-line.
\nGlobal Branch: Inspired by U-Net, this branch captures broader spatial context.
\nBoth branches are concatenated and passed through a convolution block for further processing.
\nI used a MBConvolution-like block with ECA (Efficient Channel Attention) and spatial attention.

\n

ConvLSTM
\nThe ConvLSTM layer plays a crucial role as the main autoregressive component, predicting the next frame in the form of a latent vector.

\n

Transformer
\nThe Transformer refines the latent vector by leveraging attention mechanisms, enabling the model to capture long-range temporal dependencies.

\n

Decoder
\nThe Decoder takes the refined latent vector and outputs the predicted fire-line for the next timestep.

\n

Edit:
\nTraining notebook: https://www.kaggle.com/code/jobayerhossain/flame-ai-training/notebook
\nInference notebook: https://www.kaggle.com/code/jobayerhossain/flameai-inference/notebook

", "rawMarkdown": "First of all, thank you to the hosts for organizing this competition. It was a truly great experience competing with all of you!\n\n******My Solution******\n****Feature Engineering****\n**Inclusion of Fire-Line Differences:**\nI computed the difference between the fire-line at the current timestep and the previous timestep\n\n**Inclusion of Surrounding Positions:**\nTo improve spatial awareness, I included information from surrounding positions.\n\n****Model Architecture****\nThe model architecture consists of four main parts: CNN -> ConvLSTM -> Transformer -> Decoder.\n\n**CNN**\nThe CNN handles spatial feature extraction with two branches:\n\nLocal Branch: This branch captures fine-grained, localized information from the fire-line.\nGlobal Branch: Inspired by U-Net, this branch captures broader spatial context.\nBoth branches are concatenated and passed through a convolution block for further processing.\nI used a MBConvolution-like block with ECA (Efficient Channel Attention) and spatial attention.\n\n**ConvLSTM**\nThe ConvLSTM layer plays a crucial role as the main autoregressive component, predicting the next frame in the form of a latent vector.\n\n**Transformer**\nThe Transformer refines the latent vector by leveraging attention mechanisms, enabling the model to capture long-range temporal dependencies.\n\n**Decoder**\nThe Decoder takes the refined latent vector and outputs the predicted fire-line for the next timestep.\n\nEdit:\nTraining notebook: https://www.kaggle.com/code/jobayerhossain/flame-ai-training/notebook\nInference notebook: https://www.kaggle.com/code/jobayerhossain/flameai-inference/notebook", "votes": 7 }, { "id": 3076723, "postDate": "2024-12-20T08:05:53.807Z", "content": "

Many Congratulations and thank you for sharing your work!

", "rawMarkdown": "Many Congratulations and thank you for sharing your work!" } ], "complete": true, "source": "live", "fetched_at": "2026-10-19T13:44:18Z", "competition_id": 85210, "competition_slug": "2024-flame-ai-challenge", "team_id": 12694063, "solution_url": "/competitions/2024-flame-ai-challenge/writeups/jobayer-hossain-flame-ai-second-place-solution" }