rohitg00--ai-engineering-from-scratch
wehub-resource-sync
65 行
4.2 KiB
JSON
65 行
4.2 KiB
JSON
{
|
|
"questions": [
|
|
{
|
|
"stage": "pre",
|
|
"question": "What is Promptable Concept Segmentation (PCS), introduced by SAM 3?",
|
|
"options": [
|
|
"Segmenting all instances of a concept described by a short noun phrase or image exemplar in a single forward pass, returning masks plus unique instance IDs",
|
|
"A visual prompt (point/box) segmentation like classic SAM",
|
|
"A multi-step chain of a detector followed by SAM",
|
|
"A tracking-only mode"
|
|
],
|
|
"correct": 0,
|
|
"explanation": "PCS is SAM 3's signature capability. You pass 'yellow school bus' or an image exemplar of a bus; SAM 3 returns every matching instance with its own mask and unique ID, plus a presence score. Classic SAM needed one prompt per instance; PCS produces all matches end-to-end."
|
|
},
|
|
{
|
|
"stage": "pre",
|
|
"question": "Why keep a decoupled detector + SAM 2 pipeline (Grounded SAM 2) in 2026 if SAM 3 already does text-prompted segmentation?",
|
|
"options": [
|
|
"Grounded SAM 2 is always more accurate",
|
|
"Modularity \u2014 you can swap in different open-vocabulary detectors (DINO-X, Florence-2, Grounding DINO 1.5) for different domains, license constraints, or threshold behaviour; SAM 3's architecture is monolithic",
|
|
"SAM 3 does not support video",
|
|
"SAM 3 is only available in the cloud"
|
|
],
|
|
"correct": 1,
|
|
"explanation": "Grounded SAM 2 is a composition of a detector and SAM 2 with frozen weights. That modularity is sometimes exactly what you need \u2014 a medical-imaging detector, a license-friendly detector, or tight threshold control. SAM 3 is more accurate end-to-end for common use but harder to customise. Both have production roles in 2026."
|
|
},
|
|
{
|
|
"stage": "post",
|
|
"question": "SAM 3's presence head produces what?",
|
|
"options": [
|
|
"The list of candidate bounding boxes",
|
|
"A tracking memory bank",
|
|
"A scalar probability that the queried concept exists in the image, decoupled from localisation; lets the model say 'not present' cleanly and reduces false positives on absent concepts",
|
|
"The final mask"
|
|
],
|
|
"correct": 2,
|
|
"explanation": "The presence head separates the 'is this here?' decision from the 'where is it?' decision. A model that must produce boxes whenever asked tends to hallucinate matches for absent concepts. With a presence head, SAM 3 can return zero detections cleanly. This also improves discrimination between closely related prompts (e.g., 'a player in white' vs 'a player in red')."
|
|
},
|
|
{
|
|
"stage": "post",
|
|
"question": "SAM 3.1 Object Multiplex (March 2026) introduced a shared-memory mechanism for tracking. What does it replace?",
|
|
"options": [
|
|
"SAM 2 altogether",
|
|
"Per-instance separate memory banks; Multiplex collapses them into one shared memory with per-instance queries so tracking N objects runs substantially faster while keeping accuracy",
|
|
"The whole presence head",
|
|
"The ViT backbone"
|
|
],
|
|
"correct": 1,
|
|
"explanation": "Prior SAM 2 / SAM 3 tracking maintained one memory bank per tracked instance, so cost grew linearly with object count. Object Multiplex introduces a single shared memory plus per-instance queries that fetch instance-specific features. Many-instance tracking is now efficient \u2014 essential for crowds and dense multi-object scenes."
|
|
},
|
|
{
|
|
"stage": "post",
|
|
"question": "You need real-time open-vocabulary DETECTION (boxes only, no masks) on an edge device. Which 2026 model family is the right choice?",
|
|
"options": [
|
|
"SAM 3",
|
|
"A custom CLIP + detector chain",
|
|
"Grounded SAM 2",
|
|
"YOLO-World (and related real-time open-vocab detectors); SAM 3 produces masks and is heavier, while YOLO-World is designed specifically for boxes at high fps on edge hardware"
|
|
],
|
|
"correct": 3,
|
|
"explanation": "YOLO-World and similar models (OV-DINO, LLMDet) are real-time open-vocabulary detectors. They do not produce masks but they hit 30-60 fps at 640x640 on modest GPUs. SAM 3 is the right choice when masks and tracking matter. Always match tool weight to the task: detection-only is fine without mask overhead."
|
|
}
|
|
]
|
|
}
|