Mismatch between the expected and actual view names for the Argoverse subset
Argoverse samples should include the following camera views:
['Front_Left', 'Front', 'Front_Right', 'Right', 'Back_Right', 'Back_Left', 'Left']
while the samples currently uploaded on HF contain keys such as:
{'Back', 'Back_Left', 'Back_Right', 'Front', 'Front_Left', 'Front_Right', 'Side_Left', 'Side_Right'}
For example, here is one actual entry from the dataset:
{'idx': 110,
'source': 'argoverse',
'category': 'Travel_Time',
'images': {'Back': '',
'Back_Left': 'ring_rear_left_315974070110418336.jpg',
'Back_Right': 'ring_rear_right_315974070110429216.jpg',
'Front': 'ring_front_center_315974070110418712.jpg',
'Front_Left': 'ring_front_left_315974070110424088.jpg',
'Front_Right': 'ring_front_right_315974070110423480.jpg',
'Side_Left': 'ring_side_left_315974070110424584.jpg',
'Side_Right': 'ring_side_right_315974070110424656.jpg'},
'question': 'These are seven camera views mounted on an ego car\n\nFront Left view: <image>\nFront view: <image>\nFront Right view: <image>\nRight view: <image>\nBack Right view: <image>\nBack Left view: <image>\nLeft view: <image>\nHow much time does it take to move from the location of the silver sedan car in the front left view directly towards to the location of the white utility trailer behind a white van in the back left view with the speed of 2 m/s?',
'options': ['A.Less than 3 seconds',
'B.3-5 seconds',
'C.5-10 seconds',
'D.More than 10 seconds'],
'answer': 'C'}
Argoverse provides 7 camera views, and we include all seven in each sample. However, datasets like Argoverse, Waymo, and NuScenes differ in both the number of cameras and the naming conventions for their views.
To standardize across datasets, we use the following unified set of camera names:
{'Front', 'Front_Left', 'Front_Right', 'Back', 'Back_Left', 'Back_Right', 'Side_Left', 'Side_Right'}
If a dataset doesn’t include one or more of these views (for example, Argoverse doesn’t have the 'Back' view), we simply leave those entries empty.
Since the 'image' field is passed as a dictionary, the order of views does not matter. That means, in the inference code you need to make sure that the order of images are correct using 'keys' in 'image'
You can try running inference using our code here:
🔗 https://github.com/vbdi/Ego3D-Bench
Please let me know if this clarifies your issue or if you’re still encountering any problems.
As you can see in the above example, Argoverse samples require the "Right" and "Left" views, but those views are not provided.
Can I just use "Side_Right" and "Side_Left" for "Right" and "Left", respectively?
Yes, "Right" corresponds to "Side_Right" and "Left" corresponds to "Side_Left"
I’ve corrected this line for Argoverse in our repo— thank you for pointing it out! It was a typo introduced when we renamed the camera views:
https://github.com/vbdi/Ego3D-Bench/blob/9bf53d0cbdacd05c248fbe294b43dc604a71040a/models/qwen2.5_vl.py#L87
https://github.com/vbdi/Ego3D-Bench/blob/9bf53d0cbdacd05c248fbe294b43dc604a71040a/models/internvl3.py#L90