Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -9,7 +9,6 @@ tags:
|
|
| 9 |
- vitra
|
| 10 |
- mano
|
| 11 |
- hand-object-interaction
|
| 12 |
-
- egocentric
|
| 13 |
---
|
| 14 |
|
| 15 |
# taco-annotations-v1
|
|
@@ -20,8 +19,8 @@ VITRA-style hand episodes for **TACO**, with per-hand instructions and paraphras
|
|
| 20 |
|
| 21 |
| | |
|
| 22 |
|---|---|
|
| 23 |
-
| episodes | **
|
| 24 |
-
| training samples (`index_frame_pair` rows) | **
|
| 25 |
| annotation | MANO pose + world/camera joints + per-frame extrinsics |
|
| 26 |
| text | one instruction per episode + 1.96 paraphrases on average |
|
| 27 |
| **images / video** | **not included** — see *Getting the frames* below |
|
|
@@ -43,6 +42,9 @@ round 2 checks the sentence belongs to that hand, strips same-hand references
|
|
| 43 |
("Rinse the right hand." -> "Rinse the hand.", because training already prepends
|
| 44 |
`Left hand: ... Right hand: ...`), and writes 1-3 paraphrases.
|
| 45 |
|
|
|
|
|
|
|
|
|
|
| 46 |
|
| 47 |
### Files
|
| 48 |
|
|
@@ -55,7 +57,7 @@ episode_frame_index.npz index_frame_pair (N,2) uint32 + index_to_episode_
|
|
| 55 |
that episode). `len(index_frame_pair)` is the size of the training set.
|
| 56 |
|
| 57 |
```python
|
| 58 |
-
import numpy as np
|
| 59 |
# tar -xf taco.tar
|
| 60 |
z = np.load("episode_frame_index.npz", allow_pickle=True)
|
| 61 |
ep_slot, frame_id = z["index_frame_pair"][sample_id]
|
|
@@ -78,9 +80,7 @@ sequential counter drifts silently if the decoder ever skips a frame).
|
|
| 78 |
|
| 79 |
### Known limitations
|
| 80 |
|
| 81 |
-
- Episodes whose round-1 caption came back `N/A` have an empty `text`; they are still in the
|
| 82 |
-
index. Filter on `text` if you only want captioned episodes.
|
| 83 |
- Paraphrase count averages 1.96, not a fixed number. Past 3 the model starts inventing;
|
| 84 |
a sentence with no prepositional phrase honestly supports only one or two.
|
| 85 |
-
- Verified:
|
| 86 |
-
points at a missing episode.
|
|
|
|
| 9 |
- vitra
|
| 10 |
- mano
|
| 11 |
- hand-object-interaction
|
|
|
|
| 12 |
---
|
| 13 |
|
| 14 |
# taco-annotations-v1
|
|
|
|
| 19 |
|
| 20 |
| | |
|
| 21 |
|---|---|
|
| 22 |
+
| episodes | **23,757** |
|
| 23 |
+
| training samples (`index_frame_pair` rows) | **736,136** |
|
| 24 |
| annotation | MANO pose + world/camera joints + per-frame extrinsics |
|
| 25 |
| text | one instruction per episode + 1.96 paraphrases on average |
|
| 26 |
| **images / video** | **not included** — see *Getting the frames* below |
|
|
|
|
| 42 |
("Rinse the right hand." -> "Rinse the hand.", because training already prepends
|
| 43 |
`Left hand: ... Right hand: ...`), and writes 1-3 paraphrases.
|
| 44 |
|
| 45 |
+
**Episodes with no instruction are not included.** Round 1 returns `N/A` when an episode
|
| 46 |
+
shows no object interaction. Those episodes are excluded from both the archive and the
|
| 47 |
+
index, so every episode here has a usable instruction.
|
| 48 |
|
| 49 |
### Files
|
| 50 |
|
|
|
|
| 57 |
that episode). `len(index_frame_pair)` is the size of the training set.
|
| 58 |
|
| 59 |
```python
|
| 60 |
+
import numpy as np
|
| 61 |
# tar -xf taco.tar
|
| 62 |
z = np.load("episode_frame_index.npz", allow_pickle=True)
|
| 63 |
ep_slot, frame_id = z["index_frame_pair"][sample_id]
|
|
|
|
| 80 |
|
| 81 |
### Known limitations
|
| 82 |
|
|
|
|
|
|
|
| 83 |
- Paraphrase count averages 1.96, not a fixed number. Past 3 the model starts inventing;
|
| 84 |
a sentence with no prepositional phrase honestly supports only one or two.
|
| 85 |
+
- Verified: the index lists exactly the episodes that have an instruction, every episode's
|
| 86 |
+
stored frame count matches its index rows, and no index entry points at a missing episode.
|