Struct Sam2VideoState 

pub struct Sam2VideoState {
    pub memory: Vec<Sam2MemoryEncoderOutput>,
    pub obj_ptr_queue: Vec<Vec<f32>>,
}

Per-track state for Sam2::predict_video_frame. Stores up to max_obj_ptrs_in_encoder past memory tokens + the rolling object-pointer queue.

Fields

memory: Vec<Sam2MemoryEncoderOutput>

Each entry: (features [out_dim, h, w] flat, pos [..., h, w] flat, h, w).

obj_ptr_queue: Vec<Vec<f32>>

Implementations

impl Sam2VideoState

pub fn new() -> Sam2VideoState

pub fn num_obj_ptr_tokens(&self, _mem_dim: usize) -> usize

Total number of memory tokens (spatial + obj-ptr) in the concatenated memory bank. mem_dim is the obj-pointer channel dim (typically 64).

pub fn assembled_memory( &self, kv_in_dim: usize, _mem_dim: usize, ) -> (Vec<f32>, Vec<f32>, usize)

Concatenate the per-frame memories into a single (memory [N_mem, kv_in_dim], memory_pos [N_mem, kv_in_dim]) pair for the memory-attention call. Spatial tokens go first, object-pointer tokens at the tail (so num_k_exclude_rope works correctly).

Trait Implementations

impl Default for Sam2VideoState

fn default() -> Sam2VideoState

Returns the “default value” for a type.