| --- |
| license: mit |
| library_name: onnx |
| pipeline_tag: video-classification |
| tags: |
| - transnetv2 |
| - shot-boundary-detection |
| - scene-detection |
| - video |
| - onnx |
| --- |
| |
| # TransNetV2 (ONNX) |
|
|
| `transnetv2.onnx` — an ONNX export of **TransNetV2** for shot boundary |
| detection (finding the hard cuts and gradual transitions between the shots of |
| a video). Runs client-side with onnxruntime-web. |
|
|
| ## Inputs / outputs |
|
|
| - **Input:** `input` float32 `[1, 100, 27, 48, 3]` — a window of 100 frames, |
| each resized to 48x27, RGB, values 0..255 (the graph divides by 255 itself). |
| - **Outputs:** two float32 `[1, 100, 1]` per-frame **probabilities** (the |
| sigmoid is already applied): |
| - single-frame — the frame a cut lands on (the last frame of the outgoing shot); |
| - many-hot — every frame inside a transition (dissolve, fade, …). |
| - Upstream inference slides the window by 50 frames with 25 frames of context |
| either side and keeps the middle 50 predictions of each window. |
|
|
| ## Provenance & license |
|
|
| - **Model:** TransNetV2 — *TransNet V2: An effective deep network architecture |
| for fast shot transition detection* (Souček & Lokoč, 2020), |
| [soCzech/TransNetV2](https://github.com/soCzech/TransNetV2). Code and model |
| weights are released under the **MIT license** |
| (`Copyright (c) 2020 Tomáš Souček`), included here as `LICENSE`. |
| - **ONNX export:** [elya5/transnetv2](https://huggingface.co/elya5/transnetv2) |
| (MIT), mirrored verbatim. |
|
|
