Model
- Base model:
Qwen/Qwen3-VL-8B-Instruct
- Adapter type: LoRA
- Training framework: ms-swift
- Main task: accident center localization
- Coordinate format:
relative_1000_xy
- Output format:
{"accident_center":[523,417]}
Task description
Given a traffic accident video clip and an accident time condition, the model predicts the center point of the accident region.
The coordinate system follows the relative_1000_xy format:
[0, 0] is the top-left corner of the image.[1000, 1000] is the bottom-right corner of the image.accident_center is the predicted center point of the accident region.
Example prompt
<video>
The accident occurs at approximately 3.70 seconds in this clip.
Locate the center point of the accident region at this moment.
Use coordinates where [0,0] is the top-left corner and [1000,1000] is the bottom-right corner of the image.
Output JSON only with key: accident_center.
Expected output:
{"accident_center":[523,417]}
Inference example
export IMAGE_MAX_TOKEN_NUM=1024
export VIDEO_MAX_TOKEN_NUM=1024
export FPS_MAX_FRAMES=48
export FPS=4
export FORCE_QWENVL_VIDEO_READER=torchcodec
swift infer \
--model Qwen/Qwen3-VL-8B-Instruct \
--adapters yzk9/accident-vlm-s8c-stageb-qwen3vl8b-lora \
--val_dataset path/to/spatial_infer.jsonl \
--max_new_tokens 64 \
--temperature 0 \
--result_path spatial_results.jsonl
Notes
- This repository only contains the LoRA adapter weights, not the base model weights.
- The adapter is intended for ACCIDENT-style traffic accident video understanding.
- The model expects an accident time condition in the prompt.
- For oracle-time evaluation, the accident time comes from ground-truth annotation.
- For end-to-end evaluation, the accident time can be provided by a temporal grounding model such as T8-SAPO.
- The output should be parsed as JSON and evaluated in
relative_1000_xy coordinates.
Training summary
This adapter corresponds to the S8c Stage B spatial grounding model.
- Input: local accident clip with accident time condition
- Output: accident center point
- Initialization: continued from the previous spatial clip model
- Additional training: type-aware auxiliary continuation
- Video token setting:
VIDEO_MAX_TOKEN_NUM=1024
- Frame setting:
FPS=4, FPS_MAX_FRAMES=48